Ecology

Terraces are a land type that is defined by its shape. They have a distinct morphological structure and edge features that distinguish them from other land types. In this study, we define terraces as agricultural land with strip or wavy sections built on slopes greater than 2° along the contour direction. Figure 1 depicts Google Maps satellite images of terraces in the Loess Plateau region. Terraces can be distinguished from other features in remote sensing images based on their colour, morphology, texture, and structure. Terraces can be distinguished from construction land, water, glaciers, and deserts by their colours. Figure 1b–d shows terraces that are primarily green and yellow. Furthermore, terraces are generally distributed along the contour direction, and can therefore be identified based on their morphology. Terraced field ridges curve downward and resemble strips in Fig. 1b,d or circles or ovals in Fig. 1c rather than a neat grid-like distribution. These features differ in morphology from the flat land shown in Fig. 1h. Based on texture and structure, the field area of terraces can be identified based on their strong edge features, as shown in Fig. 1b–d. The edges of terraces have dark stripes caused by oblique illumination received from the sun, and the field ridge of terraces often intercepts part of the sunlight due to their height. Sloping cultivated land, as shown in Fig. 1g, has no evident terraced wall. The outline of sloping cultivated land in the high-resolution image is curved, with no prominent edge features. These findings are critical differences distinguishing terraces and sloping land in high-resolution images.Fig. 1The spatial location of the Loess Plateau and images of various types of cultivated land. (a) The spatial location of the Loess Plateau and Spatial distribution of various cultivated land types images, (b) wide strip-mounted terraces in Longxi, (c) circular wide terraces in central Yulin, (d) high resolution image of Zhuanglang County in July 2019, (e) Zhuanglang County in February 2020, (f) narrow terraces in Shangbao, Chongyi, Jiangxi Province, (g) sloping cropland in Zhenjiang Town, Laibin, Guangxi, and (h) horizontal cropland in the North China Plain.Full size imageDeep learning-based terrace extraction modelThe DLTEM is a terrace extraction model that uses deep learning algorithms and other supplementary information. Initially, a preliminary terrace distribution map was obtained using a deep learning algorithm. It was then combined with the spectral and digital elevation model (DEM) elevation information to fine-tune the results. The final spatial distribution of the terraces was produced by manual correction (Fig. 2). Traditional land classification models or methods typically superimpose spectral, elevation, and morphological texture information from remote sensing images together for training, such as random forest, which is easily ignored in training since morphological texture information accounts for a relatively small amount of the total information. This leads to significant errors while identifying land classes with textural characteristics. In contrast, the DLTEM focuses on morphological texture information from remote sensing images and classifies it into land classes, followed by auxiliary correction through additional information. Thus, this method is more suitable to extract terraces enriched with texture structure information.Fig. 2Flow chart of the deep learning-based terrace extraction model.Full size imageThe UNet++ network is a classic deep learning algorithm that is uniquely unrivaled in extracting colour, morphology, texture, and structure features from images and applying them for classification. In comparison with other Convolutional Neural Network (CNN) classification models (e.g., Fully Convolutional Networks (FCN)), it has high classification accuracy, fast computation speed, strong robustness, and provides variable importance metrics. Therefore, in this study, the UNet++ network was adopted as the network framework for deep learning; the primary data source used was high-resolution satellite imagery from 2019. DEM (SRTM v4.1) data were used to obtain the elevation information and GlobeLand30 data were used to obtain the spectral information. The results were corrected to construct the final map of the distribution of terraces in the Loess Plateau.Study areaThe Loess Plateau, one of China’s four major plateaus, is located in northern central China (34°–40° N and 103°–114° E) (Fig. 1). It is covered by a thick loess layer that ranges in thickness from 50 to 80 m, and is the world’s largest loess deposition area, covering 648,700 km2. The altitude of the Loess Plateau ranges from 800 to 3,000 m, its average annual temperature is 6–14 °C, and its average annual precipitation is 200–700 mm. Since ancient times, the Loess Plateau has been used for agriculture because of its fine grains, fluffy soil texture, and rich soluble mineral nutrients, all of which are conducive to crop cultivation. However, long-term unsustainable land use caused the degradation of the vegetation cover in the Loess Plateau. Moreover, the land is degrading due to considerable nutrient loss caused by long-term water erosion in conjunction with natural conditions, such as arid climate, loose soil, concentrated and heavy rainfall. The fragmented ground in the region has made it susceptible to soil erosion. It has also become the primary source of Yellow River sediment as a result of the massive flow of eroded sediment into the Yellow River, posing a serious threat to the economic and social development of the lower Yellow River basin.Terracing is one of the main measures used to enhance crop yield and conserve soil and water in the region. Since the 1980s, the Chinese government has implemented many large-scale slope-to-terrace projects in the Loess Plateau. Especially in recent years, the outline of the comprehensive management plan for the Loess Plateau area (2010–2030) has been promulgated with a planned area of 2.608 million hectares for slope to terrace conversion, making it the core area of slope to terrace conversion projects in the country.Data preparationAlthough high-resolution satellite images can be an important data source for the spatial distribution of terraces on the Loess Plateau, they are not ideal for terraces classification. On the one hand, a higher resolution image requires more storage space. On the other hand, it reduces the efficiency, prolongs the interpretation time, and increases the noise in the image, affecting the interpretation accuracy. Most of the terraces on the Loess Plateau are wider than 7 m (Fig. 1b–d). These are wide terraces in comparison with the narrow terraces of southern China (Fig. 1f), which are less than 2 m wide. Furthermore, it is also easy to mistake the fish-scale pits constructed for soil and water conservation for terraces because of their similarity in form. However, as the width of their field surface is less than 1.5 m, remote sensing images with a 2 m resolution can effectively prevent the false extraction of such features. Based on the actual situation of this study area, we chose a high-resolution image with a spatial resolution of 1.89 m from Google Maps 16 level as the data source. The colour, texture, and morphological features of terraces in the images show seasonal variations. In autumn and winter, the weather is dry, and the vegetation is less shaded in the Loess Plateau. During this time, even the edge features become more visible and easier to identify. As a result, we selected images from October 2018 to February 2019 whenever possible (Fig. 1c,d).Deep learning network selectionLand classification is the extraction of land types from remote sensing images using image segmentation techniques. As the key technology of image segmentation, the Fully Convolutional Network (FCN) classifies images at the pixel level. FCN follows the network structure pattern of encoding and decoding, which adopts AlexNet as the encoder of the network and then employs transposed convolution to up-sample the feature map output from the final convolutional layer of the encoder to the resolution of the input image to achieve pixel-level image segmentation. However, due to the large error in image pixel boundary localization, Ronneberger et al.29 improved the FCN structure in 2015 by expanding the capacity of the network decoder by adding a contracting path to the encoding and decoding modules to achieve more accurate pixel boundary localisation29. The U-Net network is commonly used in medical image processing because it requires a small number of training samples and is effective in classifying objects with a fixed structure and limited semantic information. This network is comparable to natural image semantic segmentation such as Deeplab v3+, which has a smaller number of model parameters and the same effect.Since the texture and morphological features of terraces and human organs have certain similarities, they are primarily manifested by simple semantic information contained within the terrace images themselves. Thus, high-level semantic information and low-level features of such images become more important. However, high-resolution images are more complicated and variable than medical image patterns, and errors in terrace extraction edge identification using the U-Net network, such as boundary segmentation of terraces and flatlands, still occur. To fully utilize the semantic information of the network, we adopted a nested U-Net architecture, namely the UNet++ network proposed by Zhou et al.28. The network integrates long-connected and short-connected architectures to capture features at different levels by adding a shallower U-Net structure and integrates them via feature superposition to make the scale difference of feature maps smaller when fused to enhance the correct rate of image segmentation edges. However, because the U-Net++ network increases the number of model parameters, this study adopted the sparse matrix approach to accelerate model training and decrease the number of parameters.Data pre-processingData pre-processing is a prerequisite for UNet++ network training, that is, valid input according to the standard format annotation before training can be performed. Since the UNet++ network proposed by Zhou et al.28. is primarily used for medical images, which have characteristics such as fixed image structure, no spatial information, and less pattern variation, labelling medical images is comparatively easier using this method. In contrast, high-resolution remote sensing images have a large number of rasters, many pattern changes, irregular image structure, and spatial information. Therefore, determining how to better annotate high-resolution remote sensing images and reduce the annotation workload becomes critical. First, we vectorized the training sample area and generated the terrace vector dataset using ArcGIS with a high-resolution remote sensing image as the primitive map. Second, we converted the terrace vector dataset into raster data. The information of the raster had to be identical to that of the primitive map, including the size of the raster, its processing range, and its coordinate system. The output was converted to TIFF format to complete the image annotation. Since the raster size input to UNet++ network training is a fixed size, it is much smaller than the original image. To simplify the process of inputting the original image and its annotation information, we added an image import module to DLTEM, which was a sliding window of 400*400, and read the image automatically by setting the corresponding judgement conditions. Finally, the entire high-resolution image was processed automatically into the model in accordance with the established rules for training.The goal of the data enhancement was to improve the universality and robustness of the UNet++ network training results. As mentioned above, the high-resolution images taken simultaneously often included clouds or other anomalies in some areas, as the images were stitched together using multiple sources of data fusion. This can easily form evident stitching traces (Fig. 1c,d) due to the different shooting times and image quality of various data sources, i.e., brightness, saturation, and colour contrast of the images. Thus, the model trained on the original image data has strong limitations, and in many scenes, there are notable matrix-type misclassification regions due to image differences, making extraction work challenging. Therefore, in this study, we first adjusted the brightness, grayscale, and contrast of the training data after input to enhance its colour feature recognition ability. We then altered the scaling of the image, and rotated and transformed the training image from 0° to 360° to enhance morphological feature recognition and the accuracy of the training network in terrace extraction.Parameter settingThe network parameter setting is the most critical hyperparameter for UNet++ network training. They are mainly divided into input image size, batch size, learning rate, number of iterations, objective function, gradient descent strategy, momentum, decay rate, and activation function. Among them, we set the image size to 400*400 pixels based on the actual situation of the terraced area, where the UNet++ network has four scaling times, and the image size must be a multiple of 16. The batch size primarily affects the convergence of the model. If the batch limit is set to one, the model is easily affected by the random perturbation phenomenon and cannot converge to find the optimal solution. Since the batch size is determined by the size of the video memory, the value of the batch is limited by equipment constraints. The model in this study used a 2080Ti video card with 11 GB of video memory, and the batch was set to 8. The learning rate, gradient descent strategy, and objective function play a role in whether the network can find the best classification model better and faster. The learning rate was set to 0.001 for the first 500 generations, with the goal of achieving fast convergence to the target region. The learning rate was then set to 0.0001 for 500–1,000 generations, and the model was fine-tuned by choosing a smaller learning rate to find the model with the highest classification accuracy. Adam was chosen for the gradient descent strategy. The momentum and adaptive learning rate were used to increase the convergence rate. The cross-entropy classification loss function was chosen as the objective function to improve the differentiation between terraced and non-terraced areas. Momentum, decay rate, and activation function were all adopted from the previous default settings of the UNet++ network.Data correctionIn this study, we primarily used high-resolution images from Google Earth as the data source to extract the distribution of terraces on the Loess Plateau. Because this image source only contains a large amount of texture structure information and no vegetation information, it is easy to misjudge and misclassify features with the same morphological structure and edge features, such as permanent snow and ice, water bodies, bare land, and artificial surfaces. Vegetation information was generally processed based on waveband data from multispectral/hyperspectral images. It requires topographic correction, atmospheric correction, radiometric calibration, de-clouding, and other operational processes, which are extremely sophisticated30.GlobeLand30 is a 30 m spatial resolution global surface coverage dataset developed by the National Geomatics Center of China. The most recent GlobeLand30 dataset (v2020) has been updated with data sources from 2017 to the present. Its extensive data sources enable effective reduction of the impacts of cloud cover, with an overall accuracy of 85.72%. The classification accuracy of permanent snow and ice, water bodies, bare land, and artificial surfaces of this dataset is as high as 75.79%, 84.70%, 81.76%, and 86.70%, respectively. Since the update time of v2020 data is similar to that of high-resolution images, it can be used as correction data for vegetation information31.Since the training image data are two-dimensional planar data with no elevation or slope information (Fig. 1g), certain flat fields with visible field bumps are easily misclassified as terraces. The Space Shuttle Radar Topography Mission (SRTM v4.1) DEM has a spatial resolution of 30 m and ranges from 60° N to 56° S, completely covering the Loess Plateau32,33. In this study, these data were treated as terrain correction data. The amendment standard corrects the areas that have been extracted as terraces below 2° to non-terraced areas according to the requirements of the Ministry of Natural Resources of China.The spatial resolution of our extracted terraces is 1.89 m, whereas the spatial resolution of GlobeLand30 and DEM as correction data sources is 30 m, which is difficult to meet the requirements of data processing. Hence, we up-sample the two correction data sources, and then used multi-source data fusion. First, we extracted and up-sampled the terraced areas of glaciers, rivers, and deserts from GlobeLand30 to a spatial resolution of 1.89 m. Secondly, we up-sampled the DEM to 1.89 m using spatial interpolation for its raster centre as the true value of the region and performed a slope calculation for the up-sampled DEM. Further, the spatial distribution maps of glaciers, rivers, deserts, and slope maps of the Loess Plateau with the same resolution as the spatial distribution maps of terraces were available. Finally, we superimposed these images, used the terrace range in the TDMLP as a mask, and assessed the pixels in the mask area one by one. If a pixel belonged to permanent snow and ice, a water body, bare land, or an artificial surface, or had a slope less than 2°, it was modified to the background value. Otherwise, the original value was retained.We made artificial corrections to the data based on the extracted results for the arid areas of the Loess Plateau as well as for the flatter basins, given that these areas do not feature terraces.Training and validation dataFor supervised classification, the selection of sample areas and sample features is crucial. The focus and core of any land classification work is representative and effective training sample selection. To obtain a better sample area selection, we considered the selection of sample areas from three perspectives, i.e., colour texture features, topographic features, and spatial distance of the training samples. First, the terraces in this study are in agricultural land, including cultivated land, woodland, grassland, and other types of land; thus, different types of land will present different texture details. At the same time, high-resolution images from Google Earth are mosaicked. Because of the different acquisition times, the same region and land type will have visible colour differences and stitching traces, which is more common in the Loess Plateau region. Therefore, these factors should be considered in the selection of training samples as much as possible to improve the generability of the model and the correct rate of its extraction. Second, the state of the terraces varies according to topographic features. Among them, gradient, direction, altitude, and climate are the most significant factors. Terraces can be categorised as shallow-slope or steep-slope terraces. Based on slope aspect, altitude, and climate characteristics, they can also be categorised as either easy to identify or hard to identify. Thus, the sample should be inclusive of these types of terraces. According to the first law of geography, terraces in different spatial locations have different morphologies. Therefore, the spatial location of the samples should also be at a certain distance.In summary, we selected one county in each region based on the geomorphic zoning characteristics of the Loess Plateau. In addition, we added one more in the area where the density of terraces may be higher. Finally, we selected the whole area of seven counties (Fig. 3) as the training sample area distribution, covering 2.18% of the overall Loess Plateau area. The colour morphological features, topographic features, spatial location, and imaging quality of terrace images in these regions are highly representative. This method was unique from other classification methods. Most of the traditional methods are based on the single-pixel information of feature layers such as random forests, which tend to ignore the neighbouring information around the point, and thus are subject to misclassification and under classification for land types with outstanding texture information. In our study, we adopted the visual interpretation of the whole domain, which can cover the neighbourhood information of each pixel point more comprehensively. To ensure the uniformity and correctness of visual interpretation, the terraces in the training area were visually interpreted by seven interpreters after uniform professional training. For the disputed and uncertain areas, the seven interpreters carried out interactive interpretation and scoring according to the interpretation results. Finally, two other interpretation experts made the final review and corrections. The interpretation results of the training area were re-examined and revised based on the results of the later interpretations.Fig. 3Distribution of training sample areas and validation sites in terraces on the Loess Plateau.Full size imageTo better assess and compare the validity and correctness of the terraced agricultural area datasets on the Loess Plateau in quantitatively, the validation dataset was divided into two parts: a per-pixel point-based validation set and a field validation dataset of terraces with location information. The extracted datasets were comprehensively evaluated in terms of both pixel scale and field validation.We constructed a single-pixel validation point that evaluates the TDMLP. We applied the Icosahedral Snyder Equal Area Discrete Global Grid created by ArcGIS. Based on this strategy, the study area was partitioned into 972 regions (Fig. 3). To better validate the terrace classification results (excluding non-terrace classes), we placed more validation points within the grid where the terrace distribution is more concentrated. First, we calculated the proportion of terraces in each hexagonal grid to the total area of the hexagonal grid. Second, we separated the terraces into four levels according to the proportion of terraces to the whole grid area as 0–20%, 20–50%, 50–80%, and 80–100% and the number of validation points was 10, 20, 40, and 50, respectively.Since the proportion of the extracted terraced area to the total area was only 14%, direct random point deployment would have led to fewer terraced validation sets and thus would have affected the final data evaluation. Therefore, in the deployment strategy, we ensured that the validation points distributed in the extracted terraces in each grid account for at least one-fifth of the total number of validation points, but for the grid with a smaller proportion of terraces or even 0, this practice was meaningless. Hence, we stipulated that in the grid with a proportion of terraces ≤1%, direct random scattering was to be performed. The final scattered verification points in the terraced and non-terraced areas were 5,194 and 6,226, respectively, with a ratio close to 1:1 for easy verification. The spatial distribution is shown in Fig. 3.We validated the spatial distribution map of terraces on the Loess Plateau from 14 April 2021 to 1 May 2021 and constructed a field validation dataset of terraces with location information. Considering the longitudinal, latitudinal, and vertical heterogeneities of the Loess Plateau, the verification route was divided into two sections, north to south and east to west, to more comprehensively cover all regions of the Loess Plateau. The verification route started at Hohhot in the northeast of the Loess Plateau. It passed through the Datong Basin, followed the Yellow River to the south and the Weihe Plain, and then travelled westward through Mount Liupan to the westernmost part of the Loess Plateau. The route was through 54 counties/districts in 16 cities and six provinces on the Loess Plateau, with a total distance of 3,680 km, covering 15.8% of the counties on the Loess Plateau (total of 341 counties). We also surveyed and sampled the verification points approximately every 5 km along the route and collected data from a total of 815 sample points, covering various types of terraces on the Loess Plateau. The results are shown in Fig. 3. More

A hippo swims in Colombia’s Magdalena River, near where Pablo Escobar’s compound was located.Credit: Fernando Vergara/AP/Shutterstock

Colombian environment minister Susana Muhamad has triggered fear among researchers that she will protect, rather than reduce, a growing population of invasive hippos that threaten the country’s natural ecosystems and biodiversity. Although she did not directly mention the hippos — a contentious issue in Colombia — Muhamad said during a speech in late January that her ministry would create policies that prioritize animal well-being, including the creation of a new division of animal protection.
Landmark Colombian bird study repeated to right colonial-era wrongs
The hippos escaped from drug-cartel leader Pablo Escobar’s estate after he died in 1993. Left alone, the male and three females that Escobar had illegally imported from a US zoo established themselves in Colombia’s Magdalena River and some small lakes nearby — part of the country’s main watershed. After years of breeding, the ‘cocaine hippos’ have multiplied to about 150 individuals, scientists estimate.Given that the hippos (Hippopotamus amphibius) — considered the largest invasive animal in the world — have no natural predators in Colombia and have been mating at a steady rate, their population could reach 1,500 in 16 years, according to a modelling study published in 20211. “I do not understand what the government is waiting for to act,” says Nataly Castelblanco Martínez, a Colombian conservation biologist at the Autonomous University of Quintana Roo in Chetumal, Mexico, and co-author of the study. “If we don’t do anything, 20 years from now the problem will have no solution.”Researchers have called for a strict management plan that would eventually reduce the wild population to zero, through a combination of culling some animals and capturing others, then relocating them to facilities such as zoos. But the subject of what to do with the hippos has polarized the country, with some enamoured by the animals’ charisma and value as a tourist attraction and others concerned about the threat they pose to the environment and local fishing communities.‘A bit surreal’Several studies and observations suggest how destructive it could be to allow the Colombian hippo population to explode. A 2019 paper2, for example, showed that, compared with lakes without hippos, those where the animals have taken up residence contain more nutrients and organic matter that favour the growth of cyanobacteria — aquatic microbes associated with toxic algal blooms. These blooms can reduce water quality and cause mass fish deaths, affecting local fishing communities.

A sign near Doradal, Colombia, warns passersby of the danger of invasive hippos.Credit: Juancho Torres/Anadolu Agency via Getty

Other scientists have predicted that the hippos could displace endangered species that are native to the Magdalena River, such as the Antillean manatee (Trichechus manatus manatus), by outcompeting them for food and space. They caution that traffic accidents and attacks on people caused by the hippos will become more common. And they warn that wildlife traffickers are already taking advantage of the situation by illegally selling baby hippos — a trend that could intensify.“It’s a bit surreal,” says Jorge Moreno Bernal, a vertebrate palaeontologist at the University of the North in Barranquilla, Colombia. “This is just a taste of what may come.”When Colombian authorities first recognized the speed at which the hippo population was growing, during the 2000s, they acted to reduce their numbers. But in 2009, when photos appeared online after soldiers gunned down Pepe, Escobar’s fugitive male hippo, the outcry from animal-rights activists and others plunged the environment ministry into an “institutional paralysis”, says Sebastián Restrepo Calle, an ecologist at Javeriana University in Bogotá.Researchers say that the hippos don’t belong in Colombia — they are native to sub-Saharan Africa. Simulations run by Castelblanco Martínez and her colleagues suggest that to reduce the population to zero by 2033, about 30 hippos would need to be removed from the wild population per year1. No other course of action, including sterilization or castration, would eradicate them, according to the modelling of various management scenarios, says Castelblanco Martínez.The cost of inactionThe worry now is that, instead of basing decisions on evidence and expertise in conservation, the government is listening to popular opinion, says Restrepo Calle. Neither Muhamad nor representatives of the environment ministry replied to Nature’s requests for comment.
Ancient stone tools suggest early humans dined on hippo
“Why prioritize one species over our own ecosystems?” — especially a species that isn’t native, asks Alejandra Echeverri, a Colombian conservation scientist at Stanford University in California. Along with her colleagues, Echeverri published a study last month showing that Colombia has few policies governing invasive species compared with its overall number of biodiversity policies3.Animals-rights advocates, meanwhile, argue that they aren’t ignoring environmental concerns. Luis Domingo Gómez Maldonado, an animal-rights activist and specialist in animal law at Saint Thomas University in Bogotá, says “It’s not about saving the hippos on a whim,” but rather about solving the issue while also giving the hippos justice. “My indisputable position is: let’s save as many individuals as possible, let’s do it ethically.”Researchers, too, say they have the animals’ best interests at heart. “Even if [advocates] don’t see it, we care about the hippos,” Castelblanco Martínez says. “The more time that passes, the more hippos will either have to be culled, castrated or captured.”The question is whether environmental authorities will act swiftly to draft and enforce a management plan that is both ethical and effective. Should they sit on the issue for too long, Castelblanco Martínez warns, rural communities that are most affected by the hippos might take matters into their own hands.If the government doesn’t cull them, she says, people will use shotguns to do it. More

Test organisms and exposuresIn this study, we used test organisms and reagents according to the Acute Toxicity Test Method of Daphnia magna Straus(Cladocera, Crustacea); ES 04704.1b29. Daphnia magna were fostered at the National Institute of Environmental Research and were adopted. During the test, adult female Daphnia magna over two weeks of age, cultured over several generations, were transferred to a freshly prepared container the day before the test. Daphnia magna are neonates for less than 24 h after birth29. To maintain the sensitivity of the organism, young individuals less than 24 h old that reproduced the following day were used. Individuals of a similar size were selected for the test. Daphnia magna was fed YCT, which is a mixture of green algae in Chlorella sp., yeast, Cerophy II(R), and trout chow. Sufficient amounts of prey were supplied 2 h before the test to minimize the effects of prey during the test. The test medium was prepared by dissolving KCl (8 mg/L), (text {MgSO}_4) (120 mg/L), (text {CaSO}_4 cdot 2 text {H}_2 text {O} ) (120 mg/L), and (text {NaHCO}_3) (192 mg/L) in deionized water.Automatic high-throughput Daphnia magna tracking systemTo build an automatic high-throughput Daphnia magna tracking system, we equipped the system with a video analysis algorithm as well as flow cells (Fig. 1). In the tracking system, six flow cells filled with culture medium were installed in the device. Each flow cell contained 10 Daphnia magna. Subsequently, to automatically measure the state of Daphnia magna, the six flow cells were photographed at 15 frames per second using a camera (Industrial Development Systems imaging) equipped with a CMOSIS sensor capable of infrared imaging. A red light close to the infrared spectrum was placed at the back of the flow cells for uniform illumination and to minimize stress on Daphnia magna. To capture the size and movement of the Daphnia magna as accurately as possible, the camera was set to a frame rate of 15 fps and a resolution of 2048 (times ) 1088 (2.23 MB), using a 12 mm lens. The distance between the flow cell and the camera was set to 16 cm. To measure the number of mobile Daphnia magna, their lethality, and swimming inhibition automatically and simultaneously, one camera for every two cells was used to collect the status data of Daphnia magna. For assessing ecotoxicity, the video analysis system used images obtained from the six flow cells to track each Daphnia magna and estimate key statistics such as the number of mobile individuals, average distance, and radius of activity.Figure 1New automatic high-throughput video tracking system for behavioral analysis using Daphnia magna as a model organismFull size imageThe automatic high-throughput video tracking system in the ecotoxicity measuring device was designed to continuously measure the ecotoxicity of Daphnia magna (Fig. 2). Daphnia magna moves faster at high temperatures and is less active at low temperatures. Thus, a constant temperature module that can be set to an appropriate Daphnia magna habitat temperature (20 ± 2 (^{circ })C) was added to create a suitable culture environment for Daphnia magna29. Natural pseudo-light ((lambda >590) nm, 3000 k) was installed on the upper part of the detector for proper habitat light intensity (500 Lux–1000 Lux). The size of the flow cell was set as small as possible while observing the movement of the Daphnia magna. An automatic feeding system was installed so that food could be injected during the replacement cycle. The six independent multi-flow cells were designed with an automatic dilution injection module; therefore, these flow cells were diluted to six different concentrations (100%, 50%, 25%, 12.5%, 6.25%, and 0%).Figure 2Schematic representation of the automatic high-throughput video tracking systemFull size imageAutomatic tracking algorithmThe CPU used for Daphnia magna tracking was Intel i5-9300H @ 2.40 GHz, with 8 GB of memory and Windows 10 Pro 64-bit operating system. In this experiment, the algorithms were trained using 12 Daphnia magna videos and tested using an additional four Daphnia magna videos. Subsequently, the detection and tracking methods were compared. The videos, each of which had a duration of 30 s, were captured at a rate of 15 frames per second. Generally, for long-time or real-time videos, the following factors must be considered in tracking Daphnia magna: automatic binarization between the object and background, effective classification of Daphnia magna or noise, and the speed of the algorithm. Therefore, to develop an efficient tracking algorithm, we propose the following tracking process (Fig. 3A). In this process, each frame is initially converted into an image and the background is identified from the obtained video (Fig. 3B). The background is the average of the frames over the previous 20 s, and the tracking system takes 20 s to capture the first background image. The background is subtracted from the image for object detection (Fig. 3C). The objects include Daphnia magna and noise such as droplets and sediment. The difference between the background and frame images is binarized, and each area of the binarized values is regarded as an object. Conventionally, the binarized values are manually generated using specific thresholds. In this study, the images are automatically binarized using k-means clustering to select the threshold value. After binarization, several machine learning methods are used to classify the objects as Daphnia magna or noise (Fig. 3D). For a faster tracking algorithm, we use simple machine learning methods such as random forest (RF) and support vector machine (SVM). The predicted Daphnia magna are tracked using SORT24, which is a fast and highly accurate tracking algorithm (Fig. 3E). Finally, based on the tracked results, statistics for assessing ecotoxicity, such as the number of mobile individuals, average distance, and radius of activity, are estimated to evaluate the toxicity of the aquatic environment.Figure 3Automatic Daphnia magna tracking algorithm process. (A) Overview of automatic tracking algorithm process. (B) Image extraction step. (C) Background subtraction step. (D) Daphnia magna detection step. (E) Daphnia magna tracking step.Full size imagek-means clustering for automatic background subtractionMany tracking algorithms assume that the background is fixed. With fixed backgrounds, the difference between the frame and background can be used to identify objects. However, automatically selecting the precise threshold value for image pixel binarization becomes one of the key problems in identifying objects. The proposed method applies k-means clustering to the pixel values of the subtracted image30, and the center value of each calculated cluster mean is selected as the threshold value (Fig. 4). In the k-means clustering method, grouping is repeatedly performed using the distance between data points31. For binarization, two groups are formed. Let (mu _1 (t)) be the mean of pixels less than the threshold and (mu _2(t)) be the mean of pixels greater than the threshold. At first, (mu _1(t), mu _2(t)) are randomly initialized. Subsequently, each pixel is grouped into a closer mean of each group. The above steps are repeated several times until the group experiences a few changes. Finally, the threshold is calculated as an average of the two means.Figure 4Example of automatic threshold value setting for binarization between objects and background using k-means clusteringFull size imageClassification methodsObject detection based solely on the subtraction between the background and frame images may have low accuracy. As the background in the proposed process is the average value of the frame images, noise may occur. Although this noise is removed by threshold selection in binarization, using only the threshold selection is not efficient for long or real-time videos. Therefore, additional noise must be classified and removed using machine learning models, requiring the construction of a database. In the database, the obtained objects are manually labeled as noise or Daphnia magna and are called ground truth. For classification, the resized 8 (times ) 8 image of each object is stored in the database. The resized image is transformed into a feature using the Sobel edge detection algorithm32 and entered as inputs to the classification models. In this study, classification models such as RF33 SVM34 were used.RF is a model that integrates several decision tree models35. All training data are sampled with a replacement for training each decision tree model. The decision tree model is trained to split intervals of each independent variable by minimizing the gini index (Eq. 1) or entropy index (Eq. 2). The gini index and entropy index denote the impurity within the intervals.$$begin{aligned} G= & {} 1- sum _{i=1}^{c} p_i ^2 end{aligned}$$
(1)
$$begin{aligned} E= & {} – sum _{i=1}^{c} p_i log_2 p_i end{aligned}$$
(2)
where (p_i) is a probability within i-th interval, and c is the number of intervals. For better performance, the RF selects independent variables of training data randomly. This step serves to reduce the correlation of each model. If predictions of each decision tree are uncorrelated, then the variance of an integrated prediction of models is smaller than the variance of each model. RF integrates several model predictions using the voting method. An advantage of the RF method is that it avoids overfitting because the model uses the average of many predictions.SVM is a model designed to search for a hyperplane to maximize the distance, or margin, between support vectors. The hyperplane refers to the plane that divides two different groups, and the support vector represents the closest vector to the hyperplane. Let (D=({textbf{x}}_i, y_i), i=1, ldots , n, {textbf{x}}_i in {mathbb {R}}^p, y_n in { -1,1 }) be training data. Suppose that the training data are completely separated linearly by a hyperplane; then, the hyperplane is expressed as Eq. 3.$$begin{aligned} {textbf{w}}^T {textbf{x}} + b = 0, end{aligned}$$
(3)
where ({textbf{w}}) is a weight vector of the hyperplane, and b is a bias. The weight vector is updated by minimizing Eq. 4.$$begin{aligned} L = {1 over 2} {textbf{w}}^T {textbf{w}} text { subject to } y_i ({textbf{w}}^T {textbf{x}} + b) ge 1 end{aligned}$$
(4)
We can transform Eqs. 4 to  5 by using the Lagrange multiplier method.$$begin{aligned} L^* = {1 over 2} {textbf{w}}^T {textbf{w}} – sum _{i=1}^n a_i { y_i ({textbf{w}}^T x_i + {-}) – 1 }, end{aligned}$$
(5)
where (a_i) is the Lagrange multiplier. We can efficiently solve Eq. 5 using a dual form. Furthermore, Eq. 5 can be solved in a case where it is not completely separated using a slack variable and a kernel trick can be used to estimate the nonlinear hyperplane.SORT trackerSORT, one of the frameworks for solving the multiple object tracking (MOT) problem, aims to achieve efficient real-time tracking24. The SORT method framework is created by combining the estimation step and the association step. The estimation step forecasts the next position of each predicted Daphnia magna. The association step matches the forecasting position and next true position of each predicted Daphnia magna. In the estimation step, the SORT framework uses the Kalman filter to forecast the position of the predicted Daphnia magna in the next frame. The position of each predicted Daphnia magna is expressed as Eq. 6.$$begin{aligned} {textbf{x}} = [u,v,s,r,{dot{u}}, {dot{v}}, {dot{s}}]^T end{aligned}$$
(6)
where u and v are the center positions of each predicted Daphnia magna, s is the scale size of the bounding box, and r is the aspect ratio of the bounding box. ({dot{u}}), ({dot{v}}), and ({dot{s}}) are the amounts of change in each variable. In the association step, to associate the forecasting position and true position, the framework adopts the intersection-over-union (IOU)36 as the association metric. The Hungarian algorithm is loaded into the SORT framework to perform fast and efficient Daphnia magna association prediction. In this study, a mixed metric of IOU36 and Euclidean distance37 was used instead of only the IOU that is used in SORT (Eq. 7) for more efficient association.$$begin{aligned} C_{ij} = (1-lambda ) {max_d – d_{ij} over max_d} + lambda cdot IOU_{ij} end{aligned}$$
(7)
where (d_{ij}) is the Euclidean distance between the i-th predicted Daphnia magna in the before frame and the j-th predicted Daphnia magna in the next frame, and (lambda ) is the weight of (IOU_{ij}). (IOU_{ij}) is the IOU between the i-th predicted Daphnia magna in the before-frame and the j-th predicted Daphnia magna in the next frame.MetricsThe binary confusion matrix consists of true positive (TP), true negative (TN), false positive (FP), and false negative (FN)38. TP is the number of cases where the predicted Daphnia magna matches the actual Daphnia magna, TN is the number of cases where the objects predicted as noise are actual noise, FP is the number of cases where the predicted Daphnia magna differs from the actual Daphnia magna, and FN is the number of cases where the objects predicted as noise are not actual noise. In this study, accuracy, recall, precision, and F1 scores (Eq. 8) were used as the metrics for comparing the machine learning methods.$$begin{aligned} begin{aligned} Accuracy&= {TP + FP over TP + TN + FP + FN} \ Recall&= {TP over TP + TN} \ Precision&= {TP over TP + FP} \ F1 score&= 2 times {Precision times Recall over Precision + Recall} end{aligned} end{aligned}$$
(8)
Standard MOT metrics to evaluate tracking performance include multi-object tracking accuracy (MOTA) and multi-object tracking precision (MOTP). An important task of MOT is to identify and track the same object across two frames. Identification (ID) precision (IDP), ID recall (IDR), ID F1 measure (IDF1), and ID switches (IDs) may be used as measures for evaluating the identification and tracking of the same objects39,40.Data analysisThe toxicity test using Daphnia magna was performed following the Korean official Acute Toxicity Test Method29. The test medium was prepared by dissolving KCl (8 mg/L), (text {MgSO}_4) (120 mg/L), (text {CaSO}_4 cdot 2 text {H}_2 text {O} ) (120 mg/L), and (text {NaHCO}_3) (192 mg/L) in deionized water. Considering that Daphnia magna are neonates for less than 24 h after birth29, five neonates were exposed to 50 mL of different concentrations of heavy metals such as Potassium dichromate, Copper(II) sulfate pentahydrate, and Lead(II) sulfate (6.25, 12.5, 25, 50, and 100%) and 50 mL of culture media. Potassium dichromate is a common inorganic reagent used as an oxidizing agent in chemical industries. Copper(II) sulfate pentahydrate is a trace material widely used in industrial processes and agriculture. A significant amount of copper is emitted in semiconductor manufacturing processes, which adversely impacts the aquatic ecosystem. When present as an ion in water, copper can be acutely toxic to aquatic organisms such as Daphnia magna. Lead(II) sulfate is another nonessential and nonbiodegradable heavy metal. It is highly toxic to numerous organisms even at low concentrations and can accumulate in aquatic ecosystems41. Twenty Daphnia magna (four replicates of five each) were exposed to each test solution for 24 h. The term “immobility” means that the Daphnia magna remains stationary after exposure to chemicals such as Potassium dichromate, Copper(II) sulfate pentahydrate, and Lead(II) sulfate. In this study, immobility was used as an endpoint identifier, and the number of mobile Daphnia magna were counted to evaluate the EC50 values for the samples using the ToxCalc 5.0 program (Tidepoll Software, USA).The locomotory responses of Daphnia magna were tested after 0, 12, 18, and 24 h of exposure at different concentrations. Potassium dichromate ((text {K}_2text {Cr}_2text {O}_7)) at 2 mg/L was connected to the Daphnia magna tracking system, and standard toxic substances were automatically diluted to 100%, 50%, 25%, 12.5%, and 6.25%. The automatic high-throughput Daphnia magna tracking system automatically measured the tracking results of a 1-minute-long video at hourly intervals. The average moving distance for 20 s of each Daphnia magna in each chamber was analyzed using a repeated measures ANOVA (RMANOVA). RMANOVA was used for the analysis of data obtained by repeatedly measuring the same Daphnia magna42. It analyzes the concentration effect excluding the time effect at each hour. The time effect means the change in average distance per 20 s. RMANOVA was implemented using the agricolae package of the R 4.0.4 program43. To remove the noise affecting RMANOVA, the Daphnia magna that remained stationary for 20 s or more were removed from the observations. In this study, we used the significance level at 5%. More

Réale, D. et al. Personality and the emergence of the pace-of-life syndrome concept at the population level. Philos. Trans. R. Soc. Lond. B 365, 4051–4063 (2010).Article 

Google Scholar 
Gosling, S. D. From mice to men: What can we learn about personality from animal research?. Psychol. Bull. 127, 45 (2001).Article 
CAS 

Google Scholar 
Dingemanse, N. J., Class, B. & Holtmann, B. Nonrandom mating for behavior in the wild?. Trends Ecol. Evol. 36, 177–179 (2021).Article 

Google Scholar 
Croft, D. P. et al. Behavioural trait assortment in a social network: Patterns and implications. Behav. Ecol. Sociobiol. 63, 1495–1503 (2009).Article 

Google Scholar 
Morton, F. B., Weiss, A., Buchanan-Smith, H. M. & Lee, P. C. Capuchin monkeys with similar personalities have higher-quality relationships independent of age, sex, kinship and rank. Anim. Behav. 105, 163–171 (2015).Article 

Google Scholar 
Su, X. et al. Agonistic behaviour and energy metabolism of bold and shy swimming crabs Portunus trituberculatus. J. Exp. Biol. https://doi.org/10.1242/jeb.188706 (2019).Article 

Google Scholar 
Jolles, J. W., King, A. J. & Killen, S. S. The role of individual heterogeneity in collective animal behaviour. Trends Ecol. Evol. 35, 278–291 (2020).Article 

Google Scholar 
Bell, A. M., Hankison, S. J. & Laskowski, K. L. The repeatability of behaviour: A meta-analysis. Anim. Behav. 77, 771–783 (2009).Article 

Google Scholar 
Frost, A. J., Winrow-Giffen, A., Ashley, P. J. & Sneddon, L. U. Plasticity in animal personality traits: Does prior experience alter the degree of boldness?. P. Roy. Soc. B-Biol. Sci. 274, 333–339 (2007).
Google Scholar 
Krause, J., James, R. & Croft, D. P. Personality in the context of social networks. Philos. Trans. R. Soc. Lond. B 365, 4099 (2010).Article 
CAS 

Google Scholar 
David, M., Auclair, Y. & Cézilly, F. Personality predicts social dominance in female zebra finches, Taeniopygia guttata, in a feeding context. Anim. Behav. 81, 219–224 (2011).Article 

Google Scholar 
Favati, A., Leimar, O. & Løvlie, H. Personality predicts social dominance in male domestic fowl. PLoS ONE 9, e103535 (2014).Article 
ADS 

Google Scholar 
McGhee, K. E. & Travis, J. Repeatable behavioural type and stable dominance rank in the Bluefin killifish. Anim. Behav. 79, 497–507 (2010).Article 

Google Scholar 
Krause, J., Croft, D. P. & James, R. Social network theory in the behavioural sciences: Potential applications. Behav. Ecol. Sociobiol. 62, 15–27 (2007).Article 
CAS 

Google Scholar 
Flack, J. C., Girvan, M., de Waal, F. & Krakauer, D. C. Policing stabilizes construction of social niches in primates. Nature 439, 426–429 (2006).Article 
ADS 
CAS 

Google Scholar 
Croft, D. P., James, R. & Krause, J. Exploring Animal Social Networks (Princeton University Press, 2008).Book 

Google Scholar 
Patriquin, K. J., Leonard, M. L., Broders, H. G. & Garroway, C. J. Do social networks of female northern long-eared bats vary with reproductive period and age?. Behav. Ecol. Sociobiol. 64, 899–913 (2010).Article 

Google Scholar 
Gomes, A. C. R., Beltrão, P., Boogert, N. J. & Cardoso, G. C. Familiarity, dominance, sex and season shape common waxbill social networks. Behav. Ecol. 33, 526–540 (2022).Article 

Google Scholar 
Croft, D. P., Krause, J. & James, R. Social networks in the guppy (Poecilia reticulata). P. Roy. Soc. B-Biol. Sci. 271, S516–S519 (2004).Article 

Google Scholar 
Pike, T. W., Samanta, M., Lindström, J. & Royle, N. J. Behavioural phenotype affects social interactions in an animal network. P. Roy. Soc. B-Biol. Sci. 275, 2515–2520 (2008).
Google Scholar 
Aplin, L. M. et al. Individual personalities predict social behaviour in wild networks of great tits (Parus major). Ecol. Lett. 16, 1365–1372 (2013).Article 
CAS 

Google Scholar 
Massen, J. J. & Koski, S. E. Chimps of a feather sit together: Chimpanzee friendships are based on homophily in personality. Evol. Hum. Behav. 35, 1–8 (2014).Article 

Google Scholar 
Rault, J.-L. Friends with benefits: Social support and its relevance for farm animal welfare. Appl. Anim. Behav. Sci. 136, 1–14 (2012).Article 

Google Scholar 
Schneider, G. & Krueger, K. Third-party interventions keep social partners from exchanging affiliative interactions with others. Anim. Behav. 83, 377–387 (2012).Article 

Google Scholar 
Fraser, O. N. & Bugnyar, T. Do ravens show consolation? Responses to distressed others. PLoS ONE 5, e10605 (2010).Article 
ADS 

Google Scholar 
Rose, P. & Croft, D. The potential of social network analysis as a tool for the management of zoo animals. Anim. Welf. 24, 123–138 (2015).Article 

Google Scholar 
Clark, F. E. Space to choose: network analysis of social preferences in a captive chimpanzee community, and implications for management. Am. J. Primatol. 73, 748–757 (2011).Article 

Google Scholar 
Corner, L., Pfeiffer, D. & Morris, R. Social-network analysis of Mycobacterium bovis transmission among captive brushtail possums (Trichosurus vulpecula). Prev. Vet. Med. 59, 147–167 (2003).Article 
CAS 

Google Scholar 
Hansen, H., McDonald, D. B., Groves, P., Maier, J. A. & Ben-David, M. Social networks and the formation and maintenance of river otter groups. Ethology 115, 384–396 (2009).Article 

Google Scholar 
Radosevich, L. M., Jaffe, K. E. & Minier, D. E. The utility of social network analysis for informing zoo management: Changing network dynamics of a group of captive hamadryas baboons (Papio hamadryas) following an introduction of two young males. Zoo Biol. 40, 503–516 (2021).Article 

Google Scholar 
Pacheco, X. P. & Madden, J. R. Does the social network structure of wild animal populations differ from that of animals in captivity?. Behav. Processes 190, 104446 (2021).Article 

Google Scholar 
Watters, J. V. & Powell, D. M. Measuring animal personality for use in population management in zoos: Suggested methods and rationale. Zoo Biol. 31, 1–12 (2012).Article 

Google Scholar 
Koski, S. E. Social personality traits in chimpanzees: temporal stability and structure of behaviourally assessed personality traits in three captive populations. Behav. Ecol. Sociobiol. 65, 2161–2174 (2011).Article 

Google Scholar 
Račevska, E. & Hill, C. M. Personality and social dynamics of zoo-housed western lowland gorillas (Gorilla gorilla gorilla). J. Zoo Aqua. Res. 5, 116–122 (2017).
Google Scholar 
Stoinski, T. S., Jaicks, H. F. & Drayton, L. A. Visitor effects on the behavior of captive western lowland gorillas: The importance of individual differences in examining welfare. Zoo Biol. 31, 586–599 (2012).Article 

Google Scholar 
Wielebnowski, N. C. Behavioral differences as predictors of breeding status in captive cheetahs. Zoo Biol. 18, 335–349 (1999).Article 

Google Scholar 
Barrett, L. P. et al. Personality assessment of headstart Texas horned lizards (Phrynosoma cornutum) in human care prior to release. Appl. Anim. Behav. Sci. 254, 105690 (2022).Article 

Google Scholar 
Rose, P. E., Brereton, J. E. & Croft, D. P. Measuring welfare in captive flamingos: Activity patterns and exhibit usage in zoo-housed birds. Appl. Anim. Behav. Sci. 205, 115–125 (2018).Article 

Google Scholar 
Rose, P. E. & Croft, D. P. Social bonds in a flock bird: Species differences and seasonality in social structure in captive flamingo flocks over a 12-month period. Appl. Anim. Behav. Sci. 193, 87–97 (2017).Article 

Google Scholar 
Rose, P. E. & Croft, D. P. Quantifying the social structure of a large captive flock of greater flamingos (Phoenicopterus roseus): Potential implications for management in captivity. Behav. Processes 150, 66–74 (2018).Article 

Google Scholar 
Rose, P. E., Croft, D. P. & Lee, R. A review of captive flamingo (Phoenicopteridae) welfare: A synthesis of current knowledge and future directions. Intern. Zoo Yearb. 48, 139–155 (2014).Article 

Google Scholar 
Rose, P. E. & Croft, D. P. Evaluating the social networks of four flocks of captive flamingos over a five-year period: Temporal, environmental, group and health influences on assortment. Behav. Processes 175, 104118 (2020).Article 

Google Scholar 
Munson, A. A., Jones, C., Schraft, H. & Sih, A. You’re just my type: Mate choice and behavioral types. Trends Ecol. Evol. 35, 823–833 (2020).Article 

Google Scholar 
Schuett, W., Tregenza, T. & Dall, S. R. Sexual selection and animal personality. Biol. Rev. 85, 217–246 (2010).Article 

Google Scholar 
Jackson, W. M. Why do winners keep winning?. Behav. Ecol. Sociobiol. 28, 271–276 (1991).Article 

Google Scholar 
Dammhahn, M. & Almeling, L. Is risk taking during foraging a personality trait? A field test for cross-context consistency in boldness. Anim. Behav. 84, 1131–1139 (2012).Article 

Google Scholar 
Van Oers, K., Drent, P. J., De Goede, P. & Van Noordwijk, A. J. Realized heritability and repeatability of risk-taking behaviour in relation to avian personalities. P. Roy. Soc. B-Biol. Sci. 271, 65–73 (2004).Article 

Google Scholar 
Hinton, M. G. et al. Patterns of aggression among captive American flamingos (Phoenicopterus ruber). Zoo Biol. 32, 445–453 (2013).Article 

Google Scholar 
Royer, E. A. & Anderson, M. J. Evidence of a dominance hierarchy in captive Caribbean flamingos and its relation to pair bonding and physiological measures of health. Behav. Processes 105, 60–70 (2014).Article 

Google Scholar 
Carere, C., Drent, P. J., Privitera, L., Koolhaas, J. M. & Groothuis, T. G. Personalities in great tits, Parus major: Stability and consistency. Anim. Behav. 70, 795–805 (2005).Article 

Google Scholar 
Jouventin, P., Lequette, B. & Dobson, F. S. Age-related mate choice in the wandering albatross. Anim. Behav. 57, 1099–1106 (1999).Article 
CAS 

Google Scholar 
Black, J. M. Partnerships in Birds: The Study of Monogamy (Oxford University Press, USA, 1996).
Google Scholar 
Estevez, I., Andersen, I.-L. & Nævdal, E. Group size, density and social dynamics in farm animals. Appl. Anim. Behav. Sci. 103, 185–204 (2007).Article 

Google Scholar 
Pickering, S. The comparative breeding biology of flamingos Phoenicopteridae at the Wildfowl and Wetlands Trust Centre, Slimbridge. Intern. Zoo Yearbook 31, 139–146 (1992).Article 

Google Scholar 
Whitehead, H. Analyzing Animal Societies: Quantitative Methods for Vertebrate Social Analysis (University of Chicago Press, 2008).Book 

Google Scholar 
Wilson, A. D., Krause, S., Dingemanse, N. J. & Krause, J. Network position: A key component in the characterization of social personality types. Behav. Ecol. Sociobiol. 67, 163–173 (2013).Article 

Google Scholar 
Renner, M. J. & Kelly, A. L. Behavioral decisions for managing social distance and aggression in captive polar bears (Ursus maritimus). J. Appl. Anim. Welf. Sci. 9, 233–239 (2006).Article 
CAS 

Google Scholar 
Stevens, E. F. & Pickett, C. Managing the social environments of flamingos for reproductive success. Zoo Biol. 13, 501–507 (1994).Article 

Google Scholar 
Franks, D. W., Ruxton, G. D. & James, R. Sampling animal association networks with the gambit of the group. Behav. Ecol. Sociobiol. 64, 493–503 (2010).Article 

Google Scholar 
Haddadi, H. et al. Determining association networks in social animals: Choosing spatial–temporal criteria and sampling rates. Behav. Ecol. Sociobiol. 65, 1659–1668 (2011).Article 

Google Scholar 
Whitehead, H. & Dufault, S. Techniques for analyzing vertebrate social structure using identified individuals. Adv. Stud. Behav. 28, 33–74 (1999).Article 

Google Scholar 
Borgatti, S.P., M., E., G., & C., F.L. UCINET for windows: software for social network analysis. Analytic Technologies: Harvard, MA (2002).Borgatti, S. P. NetDraw: graph visualization software (Analytic Technologies, 2002).
Google Scholar 
Bejder, L., Fletcher, D. & Bräger, S. A method for testing association patterns of social animals. Anim. Behav. 56, 719–725 (1998).Article 
CAS 

Google Scholar 
Farine, D. R. & Whitehead, H. Constructing, conducting and interpreting animal social network analysis. J. Anim. Ecol. 84, 1144–1163 (2015).Article 

Google Scholar 
Perdue, B. M., Gaalema, D. E., Martin, A. L., Dampier, S. M. & Maple, T. L. Factors affecting aggression in a captive flock of Chilean flamingos (Phoenicopterus chilensis). Zoo Biol. 30, 59–64 (2011).
Google Scholar 
IBMCorp. IBM SPSS Statistics for Windows. IBM Corp: Armonk, NY (2012).Clarke, K.R. & Gorley, R.N. PRIMER v6: User Manual/Tutorial. PRIMER-E, Plymouth. (2006).Kassambara, A. & Mundt, F. factoextra: Extract and Visualize the Results of Multivariate Data Analyses. (2020).RCoreTeam. R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. (2021).Budaev, S. V. Using principal components and factor analysis in animal behaviour research: Caveats and guidelines. Ethology 116, 472–480 (2010).Article 

Google Scholar 
Whitehead, H. SOCPROG programs: Analysing animal social structures. Behav. Ecol. Sociobiol. 63, 765–778 (2009).Article 

Google Scholar 
Whitehead, H. SOCPROG: Programs for analyzing social structure: Whitehead Lab (2019).Hanneman, R.A. & Riddle, M., Chapter 18: Some Statistical Tools. In: Introduction to Social Network Methods. (University of California, Riverside 2005). http://faculty.ucr.edu/~hanneman/.(2005) More

Global biodiversity has been dramatically declining over the last decades1,2,3,4. The current biodiversity crisis is primarily driven by human-induced factors, the most serious of which are land-use change, habitat fragmentation, and climate change5. While global public awareness of climate change matters is high6,7, public recognition of biodiversity loss has, historically, been low8. The understanding of biodiversity concepts highly varies among countries and social groups9,10,11: in Nigeria, the biodiversity concept was known of 20.5% of non-professional Nigerians (with basic education or no formal training) while among 88.8% of professionals with tertiary education, it reached 88.8%; 60% of participants in a study in Switzerland had never heard the term biodiversity and Chinese farmers in another pilot study have never heard about biodiversity. In the European Union, the global leader of the environmental movement on both the political and discursive levels12,13, in 2018, 71% of EU citizens had heard of biodiversity, but only around 41% of these knew what biodiversity meant14. This illiteracy is a significant constraint for conservation strategies because the development and success of actions to halt and reverse biodiversity loss strongly rely on public support15.If general awareness of biodiversity loss is low, knowledge about plant diversity is even lower16. Plants have traditionally been overlooked, and expressions such as “plant blindness”, defined as a human tendency to ignore plant species17, perfectly illustrate the situation in terms of plant conservation. And yet, current estimates suggest that two out of five plant species are threatened with extinction18. Moreover, plants play a crucial role in the world ecosystems by providing habitat, shelter, oxygen, and food, including for humans19. Local community support boosts the effectiveness of biodiversity conservation actions20,21,22. However, how biodiversity is perceived and the benefits it provides to local populations have a significant influence on this support23. Therefore, stopping the loss of plant biodiversity and the impact it has on ecosystem health and human well-being must also strive to raise public awareness on the importance of plant conservation24.A big challenge, however, is to engage people with conservation. Nowadays, in a world where a large part of the human population lives in urban areas, the contact of people with nature is declining. This is a trend that will be even more accentuated in the future25. Perhaps society’s interest in plants is decreasing because of limited exposure to plants in daily lives, schools, and work. However, by critically examining our roles as plant scientists and educators, we realize that there are probably things we could, and should, do differently. New strategies to connect people to nature are required to spark people’s interest in and knowledge of plants. Citizen science programs and mobile applications (apps) are noteworthy initiatives that are helping to achieve this goal.Citizen science is defined as the general public involvement in scientific research activities and currently is a mainstream approach to collect information and data on a wide range of scientific subjects26,27. The development of mobile technologies and the widespread use of smartphones have boosted citizen science and enabled the development of mobile apps, which are digital tools that integrate, in real-time, data from multiple sources28.The goal of this article is to show how citizen science and mobile apps can be used as educational tools to raise awareness about plant biodiversity and conservation among the general public. We focused on formal education activities, at the Bachelor of Science (BSc) level, that were designed to collect data on various aspects of plant community and functional ecology. We also present the outcomes of two informal education initiatives that used citizen science to gather data on the distribution of plant diversity. We discuss these activities and results in light of their potential to engage the public into biodiversity conservation, and as educational and outreach tools.Formal education: UniversityDuring the COVID-19 pandemic (2021), Ecology practical classes of the Bologna Bachelor Degree in Biology (Faculty of Sciences of the University of Lisbon) had to be adapted to remote learning. Fortunately, during the States of Emergency imposed by the Portuguese Government, citizens were allowed to take brief walks. Taking advantage of citizen’s ability to briefly travel outdoors, we created three activities for students, as alternatives to those typically carried out in the classroom/campus, which we describe below.Activity 1—Analysis of the impact of disturbance on plant diversity in grasslandsThe objective of this activity was for students to explore the impact of disturbance and site attributes (such as soil type) on the diversity of the herbaceous plant community and its associated pollinators. This was undertaken in grasslands located near their homes, within walking distance (due to COVID lockdown movement restrictions). To achieve this goal, we developed a comprehensive sampling protocol that included methods for (i) selecting and characterizing sampling sites based on the level of human perturbation, (ii) soil characterization, (iii) sampling, identifying, and registering plants using the iNaturalist/Biodiversity4All platform and Flora-on web (Box 1), and (iv) pollinator sampling (Supplementary Data 1). To ensure accurate plant and pollinators identification, all observations were verified by professors responsible for each topic.First, each student chose one sampling site and teachers, using photographs, classified all sites regarding their perturbation level (low, medium, and high). Then, using the sampling protocol, students were invited to study different aspects of their sampling site, in loco or at their homes. Soil samples were analysed using simple methods and available household instruments (such as plastic cups, kitchen scale, and oven). Students were introduced to soil biodiversity as well as soil parameters (humidity, texture, structure, infiltration and draining) during the remote classes. Plants were sampled using a home-made 1 m2 quadrat. All species within were counted and identified to the lowest taxonomic level possible, using the mentioned apps and website. Before plant sampling, students were also asked to count and identify pollinators within their quadrats (broad taxonomic groups, bees, butterflies, flies, beetles) for 5 min, again using the apps to aid identification.Following field sampling, students were asked to calculate two taxonomic indices of plant communities. These included species richness, which measures the number of different species that occur in a sample, and the Simpson Diversity Index, which evaluates the probability that two individuals randomly selected from a sample will belong to the same species. Students also calculated functional diversity indices such as Functional Richness and Functional Dissimilarity, since functional diversity explores functional differences between species and how these differences reflect and affect the interactions with the environment and with other species29. Then, students assessed the relation between these indices and perturbation level. They analysed several functional traits of plants that are likely to respond to local perturbation (e.g., height, leaf size). Finally, they attempted to relate plant indices with the occurrence of pollinators.Overall, students sampled 147 grasslands that were affected by low (n = 17); medium (n = 86) and high (n = 40) levels of perturbation, scattered across mainland Portugal (Fig. 1a). In total, 3015 observations corresponding to 543 species of plant and 88 of insects (Fig. 1b) were registered in the iNaturalist/Biodiversity4All project Ecologia2_FCUL, created specifically to record all of the diversity data associated with this activity. Other registered taxa included six species of molluscs and 13 of arachnids, and other occasional soil macrofauna.Fig. 1: Analysis of the impact of disturbance on plant diversity in grasslands.a Location of grasslands sampled; b Banner and overview of main results of the project created in the platform iNaturalist/Biodiversity4All to register the sampled species; c Boxplots include data of the taxonomic diversity indices (plant species richness and Simpson Diversity Index) of sampled grasslands at three different perturbation levels: low, medium and high. Central lines represent median values, box limits indicate the upper and lower quartiles, whiskers correspond to 1.5 × the interquartile range above and below the upper and lower quartiles and points are the outliers. Boxplots with different letters indicate statistically significant differences among perturbation levels based on multiple pairwise comparisons.Full size imageThe results showed that the number of species (richness) decreased consistently with the level of perturbation. Simpson Diversity Index values increased, indicating low diversity values in highly perturbed herbaceous plant communities (Fig. 1c). Results revealed a trend towards an increase in the proportion of species with lower stature as perturbation increased. However, with no clear relationship with either biodiversity or perturbation. Finally, results indicated no clear relation of pollinator abundance or richness with plant richness and diversity, although field records relate a lower number of pollinators as wind intensity increased. In fact, pollinator sampling is extremely weather sensitive, which may have contributed to the lack of consistent relationships between pollinator diversity and perturbation.Box 1 Citizen science platforms and apps used for formal and informal educational activitiesiNaturalist (https://www.inaturalist.org/home): is a social network of naturalists, citizen scientists, and biologists that is based on mapping and sharing biodiversity observations. They describe themselves as “an online social network of people sharing biodiversity information in order to help each other learn about nature”. iNaturalist may be accessed via website or mobile app. Records are validated by the iNaturalist community. Observations reached approximately 110 million as of July 2022. This app allows the development of both open-access and registration-restricted projects. BioDiversity4All (https://www.biodiversity4all.org/) is a Portuguese biodiversity citizen science platform created by the Biodiversity for All Association. This platform was founded in 2010 and is currently linked to the “iNaturalist” network43. All the projects presented in this article were developed on the Biodiversity4All platform.Flora-on (https://flora-on.pt/): this portal contains occurrence data of vascular plants from the Portuguese flora collected by project collaborators (over 575,000 records as of July 2022). Flora-on was created by the Botanical Society of Portugal (SPBotânica), a Portuguese association devoted to the promotion and study of botany in Portugal. Botanists and naturalists provide most of the data, but occasional contributors are welcomed. Records are supervised by the portal editors, ensuring the dataset’s quality level. The portal includes stunning images of leaves, flowers, fruits, and other plant parts for 2299 of the 3300 taxa occurring in Portugal44. Additionally, the portal includes a powerful search engine that allows geographical, morphological, and taxonomical searches.LeafBite (https://zoegp.science/leafbyte): is a free, open-source iPhone app that measures total leaf area as well as consumed leaf area when herbivory is present45.Leaf-IT is a free and simple Android app created for scientific purposes. It was designed to measure leaf area under challenging field conditions. It has simple features for area calculation and data output, and can be used for ecological research and education46.Activity 2—Leaf trait assessment of shrub and tree speciesStudents were asked to assess three leaf traits Specific leaf area (SLA), Specific leaf mass (LMA), and Leaf Water Content (LWC) of two or three shrub or tree species. Each species should ideally fall into one of three functional groups known for their water adaptations, namely Hydrophytes, Mesophytes and Xerophytes. Students were challenged to choose charismatic Mediterranean species that grew nearby, such as Olea europaea, Nerium oleander or Phillyrea angustifolia. Alternatively, they could take the “Quercus challenge”, which involved ranking the Portuguese oak species based on their drought tolerance. A detailed protocol was developed to assist students for this purpose (Supplementary Data 2). In this protocol was demonstrated how to calculate the leaf area using the LeafBite and Leaf-IT apps (Box 1).The students calculated the SLA, LMA, and LWC of a total of 104 species (Supplementary Data 3) belonging to the main functional groups under study. Regarding the “Quercus challenge”, they were able to classify the six most representative oak species in Portugal and confirm the relationship among these indices and their tolerance to drought (Fig. 2).Fig. 2: Leaf trait assessment of shrub and tree species: Quercus challenge.Classification of Portuguese oak species regarding their drought tolerance (higher tolerance, left-up, lower tolerance right-down).Full size imageOne of the students, accomplished to present his own learning experience related to these activities at the XXIII Conference of the Environmental Research Network of Portuguese-speaking Nations – REALP, under the title “Plant Ecology during Confinement – A Digital Approach”.Activity 3—Evaluating the impact on the biodiversity of lawn management at the University of Lisbon campusAlthough, after the lockdown, practical classes returned to the laboratories and the field in 2021/22, we continued to use the iNaturalist/Biodiversity4All platform and the Flora-on website for biodiversity registering and identification, because of the success of the activities, as evidenced by the positive comments we received from students.The goal of this activity was to study the impact of lawn management on plant diversity and pollination on the University of Lisbon campus. To accomplish this, the students described the herbaceous communities and pollinators on four lawns (named C8, RL, RR, and TT) that had different management practices (mowing and irrigation). A comprehensive document with sampling guidelines was developed (Supplementary Data 4).The project Ecologia 2 Relvados 2022 registered 100 plant and 17 pollinator species (Fig. 3a). Given that the sampling took place during a cold and rainy week, which limited pollinator activity, the low number of pollinators registered was expectable (Lawson and Rands 2019). Following these analyses, the TT lawn (Fig. 3b), which had low levels of mowing and no watering, showed a significantly higher value of diversity, indicating it had the best management strategy for these systems (Fig. 3c), if the goal is to increase biodiversity.Fig. 3: Evaluating the impact on the biodiversity of lawn management at the University of Lisbon campus.a Banner and overview of main results of the project Ecologia 2 Relvados created in the platform iNaturalist/BioDiversity4All to register the sampled species; b Location of the lawns sampled in the Campus of the University of Lisbon; c Boxplots include data of the taxonomic diversity indices (plant species richness and Simpson Diversity Index) of sampled grasslands. Central lines represent median values, box limits indicate the upper and lower quartiles, whiskers correspond to 1.5 × the interquartile range above and below the upper and lower quartiles and points are the outliers. Boxplots with different letters indicate statistically significant differences among lawns based on multiple pairwise comparisons.Full size imageInformal education: BioBlitzesIntense biological surveys known as “BioBlitz” are carried out to record all organisms found in certain locations, such as cities, protected areas, or even entire countries. They are being used all over the world to collect and share georeferenced biodiversity data30. We developed two Plant Bioblitzes based on the BioDiversity4All/iNaturalist and Flora-on platforms. Social media, such as Facebook, Instagram, and Twitter, were used to promote these events and engage citizens (Fig. 4). The BioBlitzes were developed by SPBotânica in collaboration with BioDiversity4All.Fig. 4: Bioblitz I & II – Flora of Portugal.Posters created for the promotion of the two Flora of Portugal Bioblitzes.Full size imageBioblitz I & II – Flora of PortugalThe celebration of Fascination of Plants Day (18th of May) served as the backdrop for the organization of two-weekend Bioblitzes: Bioblitz Flora of Portugal I and Bioblitz Flora of Portugal II.In 2021, the Bioblitz was solely focused on project members, which meant that only those who had voluntarily joined the initiative could participate. In total, the 119 project members registered 4234 observations of 890 plant species. In contrast, the 2022 Bioblitz was an open project (no registration required). In total, the 323 observers made 6547 records of 1198 species. To evaluate the impact of the Bioblitz events, we compared the data registered in BioDiverstiy4All during the weekends of both events (2021 and 2022) with (i) the data registered in the platform during the equivalent weekends of 2019 and 2020 and (ii) also during the weekends before both Bioblitzes. The number of species, observations, and observers increased significantly from 2019 to 2020, 2021, and 2022, but, when comparing values from 2020 with 2021 and 2022, this rise was only verified during the Bioblitz weekends, proving the importance of Bioblitzes in this increase (Fig. 5).Fig. 5: Number of observations, species and observers registered on the BioDiversity4All/iNaturalist platform over equivalent weekends in 2019, 2020, 2021, and 2022.Numbers for 2021 and 2022 correspond to the weekends in which Bioblitzes I & II – Flora of Portugal were conducted, as well as previous ones.Full size image More

Bell, A. M. Behavioural differences between individuals and two populations of stickleback (Gasterosteus aculeatus). J. Evol. Biol. 18, 464–473 (2005).Article 
CAS 
PubMed 

Google Scholar 
Dochtermann, N. A. & Jenkins, S. H. Behavioural syndromes in Merriam’s kangaroo rats (Dipodomys merriami): A test of competing hypotheses. Proc. R. Soc. Lond. B 274, 2343–2349 (2007).
Google Scholar 
Tremmel, M. & Müller, C. Insect personality depends on environmental conditions. Behav. Ecol. 24, 386–392 (2013).Article 

Google Scholar 
Zidar, J. et al. A comparison of animal personality and coping styles in the red junglefowl. Anim. Behav. 130, 209–220 (2017).Article 

Google Scholar 
Sih, A., Bell, A. & Johnson, J. C. Behavioral syndromes: An ecological and evolutionary overview. Trends Ecol. Evol. 19, 372–378 (2004).Article 
PubMed 

Google Scholar 
Réale, D. & Dingemanse, N. J. Personality and individual social specialization. In Social behaviour: Genes, ecology and evolution (eds Székely, T. et al.) 417–441 (Cambridge University Press, 2010).Chapter 

Google Scholar 
Dingemanse, N. J. & Dochtermann, N. A. Quantifying individual variation in behaviour. Mixed-effect modelling approaches. J. Anim. Ecol. 82, 39–54 (2013).Article 
PubMed 

Google Scholar 
Dingemanse, N. J., Kazem, A. J. N., Reale, D. & Wright, J. Behavioural reaction norms: Animal personality meets individual plasticity. Trends Ecol. Evol. 25, 81–89 (2010).Article 
PubMed 

Google Scholar 
Wolf, M., van Doorn, G. S. & Weissing, F. J. Evolutionary emergence of responsive and unresponsive personalities. Proc. Natl. Acad. Sci. USA 105, 15825–15830 (2008).Article 
ADS 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Ólafsdóttir, G. Á. & Magellan, K. Interactions between boldness, foraging performance and behavioural plasticity across social contexts. Behav. Ecol. Sociobiol. 70, 1879–1889 (2016).Article 
PubMed 
PubMed Central 

Google Scholar 
Mathot, K. J., Wright, J., Kempenaers, B. & Dingemanse, N. J. Adaptive strategies for managing uncertainty may explain personality-related differences in behavioural plasticity. Oikos 121(7), 1009–1020 (2012).Article 

Google Scholar 
Réale, D., Reader, S. M., Sol, D., McDougall, P. T. & Dingemanse, N. J. Integrating animal temperament within ecology and evolution. Biol. Rev. 82, 291–318 (2007).Article 
PubMed 

Google Scholar 
Coppens, C. M., de Boer, S. F. & Koolhaas, J. M. Coping styles and behavioural flexibility: Towards underlying mechanisms. Philos. Trans. R. Soc. B Biol. Sci. 365, 4021–4028 (2010).Article 

Google Scholar 
Benus, R. F., Daas, S. D., Koolhaas, J. M. & van Oortmerssen, G. A. Routine formation and flexibility in social and non-social behaviour of aggressive and non-aggressive male mice. Behaviour 112, 176–193 (1990).Article 

Google Scholar 
Dall, S. R., Houston, A. I. & McNamara, J. M. The behavioural ecology of personality: Consistent individual differences from an adaptive perspective. Ecol. Lett. 7, 734–739 (2004).Article 

Google Scholar 
Mitchell, D. J. & Biro, P. A. Is behavioural plasticity consistent across different environmental gradients and through time?. Proc. R. Soc. B. 284(1860), 20170893 (2017).Article 
PubMed 
PubMed Central 

Google Scholar 
Stamps, J. A. Individual differences in behavioural plasticities. Biol. Rev. 91, 534–567 (2016).Article 
PubMed 

Google Scholar 
Stamps, J. A. & Biro, P. A. Personality and individual differences in plasticity. Curr. Opin. Behav. Sci. 12, 18–23 (2016).Article 

Google Scholar 
Dingemanse, N. J., Both, C., Drent, P. J. & Tinbergen, J. M. Fitness consequences of avian personalities in a fluctuating environment. Proc. R. Soc. Lond. B 271, 847 (2004).Article 

Google Scholar 
Smith, B. R. & Blumstein, D. T. Fitness consequences of personality: a meta-analysis. Behav. Ecol. 19, 448–455 (2008).Article 

Google Scholar 
Dingemanse, N. J. & Réale, D. Natural selection and animal personality. Behaviour 142, 1159–1184 (2005).Article 

Google Scholar 
Duque-Wilckens, N., Trainor, B. C. & Marler, C. A. Aggression and territoriality. In Encyclopedia of animal behavior (ed. Choe, J. C.) 539–546 (Elsevier, 2019).Chapter 

Google Scholar 
AmphibiaWeb. AmphibiaWeb: Information on amphibian biology and conservation. Available at https://amphibiaweb.org (2022).Ringler, M. et al. Acoustic ranging in poison frogs—It is not about signal amplitude alone. Behav. Ecol. Sociobiol. 71, 114 (2017).Article 
PubMed 
PubMed Central 

Google Scholar 
Ringler, M., Ursprung, E. & Hödl, W. Site fidelity and patterns of short- and long-term movement in the brilliant-thighed poison frog Allobates femoralis (Aromobatidae). Behav. Ecol. Sociobiol. 63, 1281–1293 (2009).Article 

Google Scholar 
Ringler, M., Ringler, E., Magaña Mendoza, D. & Hödl, W. Intrusion experiments to measure territory size: Development of the method, tests through simulations, and application in the frog Allobates femoralis. PLoS ONE 6, e25844 (2011).Article 
ADS 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Ringler, E., Ringler, M., Jehle, R. & Hödl, W. The female perspective of mating in A. femoralis, a territorial frog with paternal care—A spatial and genetic analysis. PLoS ONE 7, e40237 (2012).Article 
ADS 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Ursprung, E., Ringler, M., Jehle, R. & Hödl, W. Strong male/male competition allows for nonchoosy females: High levels of polygynandry in a territorial frog with paternal care. Mol. Ecol. 20, 1759–1771 (2011).Article 
PubMed 

Google Scholar 
Pröhl, H. Territorial behavior in dendrobatid frogs. J Herpetol 39, 354–365 (2005).Article 

Google Scholar 
Peignier, M. et al. Exploring links between personality traits and their social and non-social environments in wild poison frogs. Behav. Ecol. Sociobiol. 76, 93 (2022).Article 

Google Scholar 
Chaloupka, S. et al. Repeatable territorial aggression in a Neotropical poison frog. Front. Ecol. Evol. 10, 398 (2022).Article 

Google Scholar 
Amézquita Torres, A. et al. Masking interference and the evolution of the acoustic communication system in the Amazonian dendrobatid frog Allobates femoralis. Evolution 60, 1874–1887 (2006).
Google Scholar 
Rodríguez López, C., Amézquita Torres, A., Ringler, M., Pašukonis, A. & Hödl, W. Calling amplitude flexibility and acoustic spacing in the territorial frog Allobates femoralis. Behav. Ecol. Sociobiol. 74, 1–10 (2020).
Google Scholar 
Asab. Guidelines for the treatment of animals in behavioural research and teaching. Anim. Behav. 159, 1–11 (2020).
Google Scholar 
Du Percie Sert, N. et al. The ARRIVE guidelines 2.0: Updated guidelines for reporting animal research. PLoS Biol 18, e3000410 (2020).Article 

Google Scholar 
Ringler, E., Mangione, R. & Ringler, M. Where have all the tadpoles gone? Individual genetic tracking of amphibian larvae until adulthood. Mol. Ecol. Resour. 15, 737–746 (2015).Article 
PubMed 

Google Scholar 
Ringler, M. et al. High-resolution forest mapping for behavioural studies in the Nature Reserve ‘Les Nouragues’, French Guiana. J. Maps 12, 26–32 (2016).Article 

Google Scholar 
Keith, D. A. et al. A function-based typology for Earth’s ecosystems. Nature 610, 513–518 (2022).Article 
ADS 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Kaefer, I. L., Montanarin, A., da Costa, R. S. & Lima, P. A. Temporal patterns of reproductive activity and site attachment of the brilliant-thighed frog Allobates femoralis from central Amazonia. J. Herpetol. 46, 549–554 (2012).Article 

Google Scholar 
Rasband, W. S. ImageJ (U. S. National Institutes of Health, 1997–2021).Bolger, D. T., Morrison, T. A., Vance, B., Lee, D. & Farid, H. A computer-assisted system for photographic mark–recapture analysis. Methods Ecol. Evol. 3, 813–822 (2012).Article 

Google Scholar 
Narins, P. M., Hödl, W. & Grabul, D. S. Bimodal signal requisite for agonistic behavior in a dart-poison frog, Epipedobates femoralis. Proc. Natl. Acad. Sci. USA 100, 577–580 (2003).Article 
ADS 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Gasser, H., Amézquita Torres, A. & Hödl, W. Who is calling? Intraspecific call variation in the aromobatid frog Allobates femoralis. Ethology 115, 596–607 (2009).Article 

Google Scholar 
Hödl, W. Dendrobates femoralis (Dendrobatidae): a handy fellow for frog bioacoustics in Proceedings of the 4th Ordinary General meeting of the Societas Europaea Herpetologica, (ed.van Gelder, J. J., Strijbosch, H. & Bergers, P.) (1987).Ursprung, E., Ringler, M. & Hödl, W. Phonotactic approach pattern in the neotropical frog Allobates femoralis: A spatial and temporal analysis. Behaviour 146, 153–170 (2009).Article 

Google Scholar 
Sonnleitner, R., Ringler, M., Loretto, M.-C. & Ringler, E. Experience shapes accuracy in territorial decision-making in a poison frog. Biol. Lett. 16, 20200094 (2020).Article 
PubMed 
PubMed Central 

Google Scholar 
Hödl, W. Phyllobates femoralis (Dendrobatidae): Rufverhalten und akustische Orientierung der Männchen (Freilandaufnahmen) in Bundesstaatliche Hauptstelle für Wissenschaftliche Kinematographie (1983).Tumulty, J. P. et al. Brilliant-thighed poison frogs do not use acoustic identity information to treat territorial neighbours as dear enemies. Anim. Behav. 141, 203–220 (2018).Article 

Google Scholar 
Fernandes, I. Y. et al. Unlinking the speciation steps: Geographical factors drive changes in sexual signals of an Amazonian Nurse-Frog through body size variation. Evol. Biol. 48, 81–93 (2021).Article 

Google Scholar 
Garcia, M. J. et al. Dueling frogs: do male green tree frogs (Hyla cinerea) eavesdrop on and assess nearby calling competitors?. Behav. Ecol. Sociobiol. 73(2), 1041 (2019).Article 

Google Scholar 
Gingras, B., Böckle, M., Herbst, C. T. & Fitch, W. T. Call acoustics reflect body size across four clades of anurans. J Zool 289(2), 143–150 (2013).Article 

Google Scholar 
Stoffel, M. A., Nakagawa, S. & Schielzeth, H. rptR: Repeatability estimation and variance decomposition by generalized linear mixed-effects models. Methods Ecol. Evol. 8, 1639–1644 (2017).Article 

Google Scholar 
Fox, J. et al. Package ‘sem’: Structural Equation Models. https://CRAN.R-project.org/package=sem (2022).Burnham, K. P. & Anderson, D. R. Model selection and multimodel inference: A practical information-theoretic approach 2nd edn. (Springer, 2002).MATH 

Google Scholar 
Hertel, A. G., Niemelä, P. T., Dingemanse, N. J. & Mueller, T. A guide for studying among-individual behavioral variation from movement data in the wild. Mov. Ecol. 8, 30 (2020).Article 
PubMed 
PubMed Central 

Google Scholar 
Hadfield, J. D. MCMC methods for multi-response generalized linear mixed models: The MCMCglmmR package. J. Stat. Softw. 33, 1–22 (2010).Article 

Google Scholar 
Bürkner, P.-C. brms: An R package for Bayesian multilevel models using stan. J. Stat. Softw. 80, 1–28 (2017).Article 

Google Scholar 
Gelman, A., Hwang, J. & Vehtari, A. Understanding predictive information criteria for Bayesian models. Stat. Comput. 24, 997–1016 (2014).Article 
MathSciNet 
MATH 

Google Scholar 
Whalen, A. & Hoppitt, W. J. E. Bayesian model selection with network based diffusion analysis. Front. Psychol. 7, 409 (2016).Article 
PubMed 
PubMed Central 

Google Scholar 
R Core Team. R: A language and environment for statistical computing (R Foundation for Statistical Computing, 2019).
Google Scholar 
Ryan, M. J., Bartholomew, G. A. & Rand, A. S. Energetics of reproduction in a neotropical frog, Physalaemus pustulosus. Ecology 64, 1456–1462 (1983).Article 

Google Scholar 
Taigen, T. L. & Wells, K. D. Energetics of vocalization by an anuran amphibian (Hyla versicolor). J. Comp. Physiol. 155, 163–170 (1985).Article 

Google Scholar 
Pough, F. H. & Taigen, T. L. Metabolic correlates of the foraging and social behaviour of dart-poison frogs. Anim. Behav. 39, 145–155 (1990).Article 

Google Scholar 
Bell, A. M., Hankison, S. J. & Laskowski, K. L. The repeatability of behaviour: A meta-analysis. Anim. Behav. 77, 771–783 (2009).Article 
PubMed 
PubMed Central 

Google Scholar 
Kelleher, S. R., Silla, A. J. & Byrne, P. G. Animal personality and behavioral syndromes in amphibians: A review of the evidence, experimental approaches, and implications for conservation. Behav. Ecol. Sociobiol. 72, 10539 (2018).Article 

Google Scholar 
Moser-Purdy, C., MacDougall-Shackleton, E. A. & Mennill, D. J. Enemies are not always dear: Male song sparrows adjust dear enemy effect expression in response to female fertility. Anim. Behav. 126, 17–22 (2017).Article 

Google Scholar  More

Data sourcesMODIS on the Aqua satellite provides a global coverage within 1–2 days. All images acquired by this satellite mission from January 2003 to December 2020 were used in our study to detect global coastal phytoplankton blooms, with a total of 0.76 million images. MODIS Level-1A images were downloaded from the Ocean Biology Distributed Active Archive Center (OB.DAAC) at NASA Goddard Space Flight Center (GSFC), and were subsequently processed with SeaDAS software (version 7.5) to obtain Rayleigh-corrected reflectance (Rrc (dimensionless), which was converted using the rhos (in sr−1) product (rhos × π) from SeaDAS)41, remote sensing reflectance (Rrs (sr−1)) and quality control flags (l2_flags). If a pixel was flagged by any of the following, it was then removed from phytoplankton bloom detection: straylight, cloud, land, high sunglint, high solar zenith angle and high sensor zenith angle (https://oceancolor.gsfc.nasa.gov/atbd/ocl2flags/). MODIS level-3 product for aerosol optical thicknesses (AOT) at 869 nm was also obtained from OB.DAAC NASA GSFC (version R2018.0), which was used to examine the impacts of aerosols on bloom trends.We examined the algal blooms in the EEZs of 153 ocean-bordering countries (excluding the EEZs in the Caspian Sea or around the Antarctic), 126 of which were found with at least one bloom in the past two decades. The EEZ dataset is available at https://www.marineregions.org/download_file.php?name=World_EEZ_v11_20191118.zip. The EEZs are up to 200 nautical miles (or 370 km) away from coastlines, which include all continental shelf areas and offer the majority of marine resources available for human use. Regional statistics of algal blooms were also performed for LMEs. LMEs encompass global coastal oceans and outer edges of coastal currents areas, which are defined by various distinct features of the oceans, including hydrology, productivity, bathymetry and trophically dependent populations42. Of the 66 LMEs identified globally, we excluded the Arctic and Antarctic regions and examined 54 LMEs. The boundaries of LMEs were obtained from https://www.sciencebase.gov/catalog/item/55c77722e4b08400b1fd8244.We used HAEDAT to validate our satellite-detected phytoplankton blooms in terms of presence or absence. The HAEDAT dataset (http://haedat.iode.org) is a collection of records of HAB events, maintained under the UNESCO Intergovernmental Oceanographic Commission and with data archives since 1985. For each HAB event, the HAEDAT records its bloom period (ranging from days to months) and geolocation. We merged duplicate entries when both the recorded locations and times of the HAEDAT events were very similar to one another, and a total number of 2,609 HAEDAT events were ultimately selected between 2003 and 2020.We used the ¼° resolution National Oceanic and Atmospheric Administration Optimum Interpolated SST (v. 2.1) data to examine the potential simulating effects of warming on the global phytoplankton trends. We also estimated the SST gradients following the method of Martínez-Moreno33. As detailed in ref. 33, the SST gradient can be used as a proxy for the magnitude of oceanic mesoscale currents (EKE). We used the SST gradient to explore the effects of ocean circulation dynamics on algal blooms.Fertilizer uses and aquaculture production for different countries was used to examine the potential effects of nutrient enrichment from humans on global phytoplankton bloom trends. Annual data between 2003 and 2019 on synthetic fertilizer use, including nitrogen and phosphorus, are available from https://ourworldindata.org/fertilizers. Annual aquaculture production includes cultivated fish and crustaceans in marine and inland waters, and sea tanks, and the data between 2003 and 2018 are available from https://ourworldindata.org/grapher/aquaculture-farmed-fish-production.The MEI, which combines various oceanic and atmospheric variables36, was used to examine the connections between El Niño–Southern Oscillation activities and marine phytoplankton blooms. The dataset is available from https://psl.noaa.gov/enso/mei/.Development of an automated bloom detection methodA recent study by the UNESCO Intergovernmental Oceanographic Commission revealed that globally reported HAB events have increased6. However, such an overall increasing trend was found to be highly correlated with recently intensified sampling efforts6. Once this potential bias was accounted for by examining the ratio between HAB events to the number of samplings5, there was no significant global trend in HAB incidence, though there were increases in certain regions. With synoptic, frequent, and large-scale observations, satellite remote sensing has been extensively used to monitor algal blooms in oceanic environments17,18,19. For example, chlorophyll a (Chla) concentrations, a proxy for phytoplankton biomass, has been provided as a standard product by NASA since the proof-of-concept Coastal Zone Color Scanner (1978–1986) era43,44. The current default algorithm used to retrieve Chla products is based on the high absorption of Chla at the blue band45,46, which often shows high accuracy in the clear open oceans but high uncertainties in coastal waters. This is because, in productive and dynamic coastal oceans, the absorption of Chla in the blue band can be obscured by the presence of suspended sediments and/or coloured dissolved organic matter (CDOM)47. To address this problem, various regionalized Chla algorithms have been developed48. Unfortunately, the concentrations of the water constituents (CDOM, sediment and Chla) can vary substantially across different coastal oceans. As a result, a universal Chla algorithm that can accurately estimate Chla concentrations in global coastal oceans is not currently available.Alternatively, many spectral indices have been developed to identify phytoplankton blooms instead of quantifying their bloom biomass, including the normalized fluorescence line height21 (nFLH), red tide index49 (RI), algal bloom index47 (ABI), red–blue difference (RBD)50, Karenia brevis bloom index50 (KBBI) and red tide detection index51 (RDI). In practice, the most important task for these index-based algorithms is to determine their optimal thresholds for bloom classification. However, such optimal thresholds can be regional-or image-specific20, due to the complexity of optical features in coastal waters and/or the contamination of unfavourable observational conditions (such as thick aerosols, thin clouds, and so on), making it difficult to apply spectral-index-based algorithms at a global scale.To circumvent the difficulty in determining unified thresholds for various spectral indices across global coastal oceans, an approach from a recent study to classify algal blooms in freshwater lakes52 was adopted and modified here. In that study, the remotely sensed reflectance data in three visible bands (red, green and blue) were converted into two-dimensional colour space created by the Commission Internationale del’éclairage (CIE), in which the position on the CIE chromaticity diagram represented the colour perceived by human eyes (Extended Data Fig. 1a). As the algal blooms in freshwater lakes were manifested as greenish colours, the reflectance of bloom-containing pixels was expected to be distributed in the green gamut of the CIE chromaticity diagram; the stronger the bloom, the closer the distance to the upper border of the diagram (the greener the water).Here, the colour of phytoplankton blooms in the coastal oceans can be greenish, yellowish, brownish, or even reddish53, owing to the compositions of bloom species (diatoms or dinoflagellates) and the concentrations of different water constituents. Furthermore, the Chla concentrations of the coastal blooms are typically lower than those in inland waters, thus demanding more accurate classification algorithms. Thus, the algorithm proposed by Hou et al.52 was modified when using the CIE chromaticity space for bloom detection in marine environments. Specifically, we used the following coordinate conversion formulas to obtain the xy coordinate values in the CIE colour space:$$begin{array}{c}x=X/(X+Y+Z)\ y=Y/(X+Y+Z)\ X=2.7689R+1.7517G+1.1302B\ Y=1.0000R+4.5907G+0.0601B\ Z=0.0000R+0.0565G+5.5943Bend{array}$$
(1)
where R, G and B represent the Rrc at 748 nm, 678 nm (fluorescence band) and 667 nm in the MODIS Aqua data, respectively. By contrast, the R, G and B channels used in Hou et al.52 were the red, green and blue bands. We used the fluorescence band for the G channel because, for a given region, the 678 nm signal increases monotonically with the Chla concentration for blooms of moderate intensity21, which is similar to the response of greenness to freshwater algal blooms. Thus, the converted y value in the CIE coordinate system represents the strength of the fluorescence. In practice, for pixels with phytoplankton blooms, the converted colours in the chromaticity diagram will be located within the green, yellow or orange–red gamut (see Extended Data Fig. 1a); the stronger the fluorescence signal is, the closer the distance to the upper border of the CIE diagram (larger y value). By contrast, for bloom-free pixels without a fluorescence signal, their converted xy coordinates will be located in the blue or purple gamut. Therefore, we can determine a lower boundary in the CIE two-dimensional coordinate system to separate bloom and non-bloom pixels, similar to the method proposed by Hou et al.52.We selected 53,820 bloom-containing pixels from the MODIS Rrc data as training samples to determine the boundary of the CIE colour space. These sample points were selected from nearshore waters worldwide where frequent phytoplankton blooms have been reported (Extended Data Fig. 2); the algal species included various species of dinoflagellates and diatoms20. A total of 80 images was used, which were acquired from different seasons and across various bloom magnitudes, to ensure that the samples used could almost exhaustively represent the different bloom conditions in the coastal oceans.We combined the MODIS FLHRrc (fluorescence line height based on Rrc) and enhanced red–green–blue composite (ERGB) to delineate bloom pixels manually. The FLHRrc image was calculated as:$$begin{array}{c}{{rm{FLH}}}_{{rm{Rrc}}}={R}_{{rm{rc}}678}times {F}_{678}-[{R}_{{rm{rc}}667}times {F}_{667}+({R}_{{rm{rc}}748}times {F}_{748}\ ,,-,{R}_{{rm{rc}}667}times {F}_{667})times (678-667)/(748-667)]end{array}$$
(2)
where Rrc667, Rrc678 and Rrc748 are the Rrc at 667, 678 and 748 nm, respectively, and F667, F678 and F748 are the corresponding extraterrestrial solar irradiance. ERGB composite images were generated using Rrc of three bands at 555 (R), 488 (G) and 443 nm (B). Although phytoplankton-rich and sediment-rich waters have high FLHRrc values, they appear as darkish and bright features in the ERGB images (Extended Data Fig. 3), respectively21. In fact, visual examination with fluorescence signals and ERGB has been widely accepted as a practical way to delineate coastal algal blooms on a limited number of images21,54,55. Note that the FLHRrc here was slightly different from the NASA standard nFLH product56, as the latter is generated using Rrs (corrected for both Rayleigh and aerosol scattering) instead of Rrc (with residual effects of aerosols). However, when using the NASA standard algorithm to further perform aerosol scattering correction over Rrc, 20.7% of our selected bloom-containing pixels failed to obtain valid Rrs (without retrievals or flagged as low quality), especially for those with strong blooms (see examples in Extended Data Fig. 4). Likewise, we also found various nearshore regions with invalid Rrs retrievals. By contrast, Rrc had valid data for all selected samples and showed more coverage in nearshore coastal waters. The differences between Rrs and Rrc were because the assumptions for the standard atmospheric correction algorithm do not hold for bloom pixels or nearshore waters with complex optical properties57. In fact, Rrc has been used as an alternative to Rrs in various applications in complex waters58,59.We converted the Rrc data of 53,820 selected sample pixels into the xy coordinates in the CIE colour space (Extended Data Fig. 1a). As expected, these samples of bloom-containing pixels were located in the upper half of the chromaticity diagram (the green, yellow and orange–red gamut) (Extended Data Fig. 1a). We determined the lower boundary of these sample points in the chromaticity diagram, which represents the lightest colour and thus the weakest phytoplankton blooms; any point that falls above this boundary represents stronger blooms. The method to determine the boundary was similar to Hou et al.52: we first binned the sample points according to the x value in the chromaticity diagram and estimated the 1st percentile (Q1%) of the corresponding Y for each bin; then, we fit the Q1% using two-order polynomial regression. Sensitivity analysis with Q0.3% (the three-sigma value) resulted in minor changes ( 1/3 AND y  > y2), it is classified as a ‘bloom’ pixel.Depending on the local region and application purpose, the meaning of ‘phytoplankton bloom’ may differ. Here, for a global application, the pixelwise bloom classification is based on the relationship (represented using the CIE colour space) between Rrc in the 667-, 678- and 754-nm bands derived from visual interpretation of the 80 pairs of FLHRrc and ERGB imagery. Instead of a simple threshold, we used a lower boundary of the sample points in the chromaticity diagram to define a bloom. In simple words, a pixel is classified as a bloom if its fluorescence signal is detectable (the associated xy coordinate in the CIE colour space located above the lower boundary). Histogram of the nFLH values from the 53,820 training pixels demonstrated the minimum value of ~0.02 mW cm−2 μm−1 (Extended Data Fig. 1a), which is in line with the lower-bound signal of K. brevis blooms on the West Florida shelf21,47. Note that, such a minimum nFLH is determined from the global training pixels, and it does not necessarily represent a unified lower bound for phytoplankton blooms across the entire globe, especially considering that fluorescence efficiency may be a large variable across different regions. Different regions may have different lower bounds of nFLH to define a bloom, and such variability is represented by the predefined boundary in the CIE chromaticity diagram in our study. Correspondingly, although the accuracy of Chla retrievals may have large uncertainties in coastal waters, the histogram of the 53,820 training pixels shows a lower bound of ~1 mg m−3 (Extended Data Fig. 1a). Similarly to nFLH, such a lower bound may not be applicable to all coastal regions, as different regions may have different lower bounds of Chla for bloom definition.Although the MODIS cloud (generated by SeaDAS with Rrc869 0.12) and Index2 ( More

Chazdon, R. L. et al. Carbon sequestration potential of second-growth forest regeneration in the Latin American tropics. Sci. Adv. 2, e1501639 (2016).Article 
PubMed 
PubMed Central 

Google Scholar 
IKI. The Bonn Challenge. https://www.bonnchallenge.org/ (2022).UNCCD. Land Degradation Neutrality. https://www.unccd.int/land-and-life/land-degradation-neutrality/overview (2022).Bastin, J.-F. et al. The global tree restoration potential. Science 365, 76–79 (2019).Article 
CAS 
PubMed 

Google Scholar 
Brancalion, P. H. S. et al. Global restoration opportunities in tropical rainforest landscapes. Sci. Adv. 5, eaav3223 (2019).Article 
PubMed 
PubMed Central 

Google Scholar 
Strassburg, B. B. N. et al. Global priority areas for ecosystem restoration. Nature 586, 724–729 (2020).Article 
CAS 
PubMed 

Google Scholar 
Erbaugh, J. T. et al. Global forest restoration and the importance of prioritizing local communities. Nat. Ecol. Evol. 4, 1472–1476 (2020).Article 
CAS 
PubMed 

Google Scholar 
Fleischman, F. et al. Restoration prioritization must be informed by marginalized people. Nature 607, E5–E6 (2022).Article 
CAS 
PubMed 

Google Scholar 
Chaturvedi, R. et al. Restoration Opportunities Atlas of India. www.india.restorationatlas.org/methodology (2022).McLain, R., Lawry, S., Guariguata, M. R. & Reed, J. Toward a tenure-responsive approach to forest landscape restoration: a proposed tenure diagnostic for assessing restoration opportunities. Land Use Policy 104, 103748 (2021).Article 

Google Scholar 
Binod, B., Bhattarcharjee, A. & Ishwar, N. M. Bonn Challenge and India: Progress on Restoration Efforts Across States and Landscapes (IUCN, 2018).Government of India. Aspirational Districts Phase 1 (vikaspedia, 2018).Government of India. Census of India. https://censusindia.gov.in/2011census/dchb/DCHB.html (2011).DeFries, R. et al. Land management can contribute to net zero. Science 376, 1163–1165 (2022).Article 
CAS 
PubMed 

Google Scholar 
Borah, B., Bhattacharya, A. & Ishwar, N. M. Bonn Challenge and India. Progress On Restoration Efforts Across States and Landscapes. https://www.bonnchallenge.org/pledges/india (2018).Gopalakrishna, T. et al. Existing land uses constrain climate change mitigation potential of forest restoration in India. Conserv. Lett. https://doi.org/10.1111/conl.12867 (2022).Dhyani, S. et al. Agroforestry to achieve global climate adaptation and mitigation targets: are South Asian countries sufficiently prepared? Forests 12, 303 (2021).Article 

Google Scholar 
Nerlekar, A. N. et al. Removal or utilization? Testing alternative approaches to the management of an invasive woody legume in an arid Indian grassland. Restor. Ecol. https://doi.org/10.1111/rec.13477 (2022).Coleman, E. A. et al. Limited effects of tree planting on forest canopy cover and rural livelihoods in Northern India. Nat Sustain 4, 997–1004 (2021).Article 

Google Scholar 
Ramprasad, V., Joglekar, A. & Fleischman, F. Plantations and pastoralists: afforestation activities make pastoralists in the Indian Himalaya vulnerable. Ecol. Soci. https://doi.org/10.5751/ES-11810-250401 (2020).DeFries, R. et al. Improved household living standards can restore dry tropical forests. Biotropica https://doi.org/10.1111/btp.12978 (2021).Lele, S., Khare, A. & Mokashi, S. Estimating and Mapping CFR Potential (ATREE, 2020).Agarwala, M. et al. Impact of biogas interventions on forest biomass and regeneration in southern India. Global Ecol. Conservation 11, 213–223 (2017).Article 

Google Scholar 
Menon, A. & Schmidt-Vogt, D. Effects of the COVID-19 pandemic on farmers and their responses: a study of three farming systems in Kerala. South India. Land 11, 144 (2022).
Google Scholar 
Fremout, T. et al. Diversity for Restoration (D4R): Guiding the selection of tree species and seed sources for climate‐resilient restoration of tropical forest landscapes. J. Appl. Ecol. 59, 664–679 (2022).Article 

Google Scholar 
Hughes, K. A. et al. Can restoration of the commons reduce rural vulnerability? A Quasi-experimental comparison of COVID-19 livelihood-based coping strategies among rural households in three Indian States. Int. J. Common. 16, 189 (2022).Article 

Google Scholar 
Madhusudan, M. D. & Vanak, A. Mapping the Distribution and Extent of India’s Semi-arid Open Natural Ecosystems. https://doi.org/10.1002/essoar.10507612.1 (2021).Vanak, A. T., Hiremath, A. J., Ganesh, T. & Rai, N. D. Filling in the (Forest) Blanks: the Past, Present and Future of India’s Savanna Grasslands (ATREE, 2017).Oxford Poverty & Human Development Initiative. Global Multidimensional Poverty Index 2018. The Most Detailed Picture to Date of the World’s Poorest People. https://ophi.org.uk/wp-content/uploads/G-MPI_2018_2ed_web.pdf (2018).Hijmans, R. J. et al. raster: Geographic Data Analysis and Modeling. https://rspatial.org/raster (2023).Bivand, R. et al. rgdal: Bindings for the ‘Geospatial’ Data Abstraction Library. https://cran.r-project.org/web/packages/rgdal/index.html (2023).QGIS Development Team. QGIS Geographic Information System (Open Source Geospatial Foundation Project, 2022). More