Philippa Kaur

RESEARCH HIGHLIGHT
03 March 2023

Human hunt at least 19% of bat species worldwide — especially flying foxes, which can have wingspans of 1.5 metres. 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

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

All procedures involving animals were conducted in accordance with the guidelines and regulations from Institutional Animal Care and Use Committee (IACUC) of the University of Florida (protocol #201509019). Tis manuscript is reported in accordance with ARRIVE guidelines.Site descriptionThis study was carried out at the North Florida Research and Education Center, in Marianna, FL (30°46′35″N 85°14′17″W, 51 m.a.s.l). The trial was performed in two experimental years (2019 and 2020) in a greenhouse.The soil used was collected from a pasture of rhizoma peanut (Arachis glabrata Benth.) and Argentine bahiagrass (Paspalum notatum Flügge) as the main forages. Without plant and root material, only soil was placed into buckets, as described below in the bucket assemblage section. Soil was classified as Orangeburg loamy sand (fine-loamy-kaolinitic, thermic Typic Kandiudults), with a pHwater of 6.7, Mehlich-1-extratable P, K, Mg and Ca concentrations of 41, 59, 63, 368 mg kg−1, respectively. Average of minimum and maximum daily temperature and relative humidity in the greenhouse for September and November (September for beetle trial due seasonal appearance of beetles, and October and November to the Pear Millet trial) in 2019 and 2020 were 11 and 33 °C, 81%; 10 and 35 °C, 77%, respectively.Biological material determinationTo select the species of beetles, a previous dung beetle sampling was performed in the grazing experiment in the same area (grass and legume forage mixture) to determine the number of dung beetle species according to the functional groups as described by Conover et al.44. Beetles were pre-sampled from March 2017 to June 2018, where Tunnelers group were dominant and represented by Onthophagus taurus (Schreber), Digitonthophagus gazella (Fabricius), Phanaeus vindex (MacLeay), Onthophagus oklahomensis (Brown), and Euniticellus intermedius (Reiche). Other species were present but not abundant, including Aphodius psudolividus (Linnaeus), Aphodius carolinus (Linnaeus), and Canthon pilularius (Linnaeus) identified as Dweller and Roller groups, respectively. The pre-sampling indicated three species from the Tunneler group were more abundant, and thereby, were chosen to compose the experimental treatments (Fig. 4).Figure 4Most abundant dung beetle species in Marianna, FL used in the current study. Credits: Carlos C.V. García.Full size imageBeetles collection and experimental treatmentsThree species of common communal dung beetles were used: O. taurus (1), D. gazella (2), and P. vindex (3). Treatments included two treatments containing only soil and soil + dung without beetles were considered as Control 1 (T1) and Control 2 (T2), respectively. Isolated species T3 = 1, T4 = 2, T5 = 3 and their combinations T6 = 1 + 2 and T7 = 1 + 2 + 3. Dung beetles were trapped in the pasture with grazing animals using the standard cattle-dung-baited pitfall traps, as described by Bertone et al.41. To avoid losing samples due to cattle trampling, 18 traps were randomized in nine paddocks (two traps per paddock) and installed protected by metal cages, and after a 24-h period, beetles were collected, and the traps removed. Table 1 shows the number of dung beetles, their total mass (used to standardize treatments) per treatment, and the average mass per species. To keep uniformity across treatments we kept beetle biomass constant across species at roughly 1.7 to 1.8 g per assemblage (Table 1). Twenty-four hours after retrieving the beetles from the field traps, they were separated using an insect rearing cage, classified, and thereafter stored in small glass bottles provided with a stopper and linked to a mesh to keep the ventilation and maintaining the beetles alive.Table 1 Total number and biomass of dung beetles per treatment.Full size tableBuckets assemblageThe soil used in the buckets was collected from the grazing trial in two experimental years (August 2019 and August 2020) across nine paddocks (0.9 ha each). The 21 plastic buckets had a 23-cm diameter and 30-cm (0.034 m2) and each received 10 kg of soil (Fig. 5). At the bottom of the recipient, seven holes were made for water drainage using a metallic mesh with 1-mm diameter above the surface of the holes to prevent dung beetles from escaping. Water was added every four days to maintain the natural soil conditions at 60% of the soil (i.e., bucket) field capacity (measured with the soil weight and water holding capacity of the soil). Because soil from the three paddocks had a slightly different texture (sandy clay and sandy clay loam), we used them as the blocking factor.Figure 5Bucket plastic bucket details for dung beetle trial.Full size imageThe fresh dung amount used in the trial was determined based on the average area covered by dung and dung weight (0.05 to 0.09 m2 and 1.5 to 2.7 kg) from cattle in grazing systems, as suggested by Carpinelli et al.45. Fresh dung was collected from Angus steers grazing warm-season grass (bahiagrass) pastures and stored in fridge for 24 h, prior to start the experiment. A total of 16.2 kg of fresh dung was collected, in which 0.9 kg were used in each bucket. After the dung application, dung beetles were added to the bucket. To prevent dung beetles from escaping, a mobile plastic mesh with 0.5 mm diameter was placed covering the buckets before and after each evaluation. The experiment lasted for 24 days in each experimental year (2019 and 2020), with average temperature 28 °C and relative humidity of 79%, acquired information from the Florida Automated Weather Network (FAWN).Chamber measurementsThe gas fluxes from treatments were evaluated using the static chamber technique46. The chambers were circular, with a radius of 10.5 cm (0.034 m2). Chamber bases and lids were made of polyvinyl chloride (PVC), and the lid were lined with an acrylic sheet to avoid any reactions of gases of interest with chamber material (Fig. 6). The chamber lids were covered with reflective tape to provide insulation, and equipped with a rubber septum for sampling47. The lid was fitted with a 6-mm diameter, 10-cm length copper venting tube to ensure adequate air pressure inside the chamber during measurements, considering an average wind speed of 1.7 m s−148,49. During measurements, chamber lids and bases were kept sealed by fitting bicycle tire inner tubes tightly over the area separating the lid and the base. Bases of chambers were installed on top of the buckets to an 8-cm depth, with 5 cm extending above ground level. Bases were removed in the last evaluation day (24th) of each experimental year.Figure 6Static chamber details and instruments for GHG collection in the dung beetle trial.Full size imageGas fluxes measurementsThe gas fluxes were measured at 1000 h following sampling recommendations by Parkin & Venterea50, on seven occasions from August 28th to September 22nd in both years (2019 and 2020), being days 0, 1, 2, 3, 6, 12, and 24 after dung application. For each chamber, gas samples were taken using a 60-mL syringe at 15-min intervals (t0, t15, and t30). The gas was immediately flushed into pre-evacuated 30-mL glass vials equipped with a butyl rubber stopper sealed with an aluminium septum (this procedure was made twice per vial and per collection time). Time zero (t0) represented the gas collected out of the buckets (before closing the chamber). Immediately thereafter, the bucket lid was tightly closed by fitting the lid to the base with the bicycle inner tube, followed by the next sample deployment times.Gas sample analyses were conducted using a gas chromatograph (Trace 1310 Gas Chromatograph, Thermo Scientific, Waltham, MA). For N2O, an electron capture detector (350 °C) and a capillary column (J&W GC packed column in stainless steel tubing, length 6.56 ft (2 M), 1/8 in. OD, 2 mm ID, Hayesep D packing, mesh size 80/100, pre-conditioned, Agilent Technologies) were used. Temperature of the injector and columns were 80 and 200 °C, respectively. Daily flux of N2O-N (g ha−1 day−1) was calculated as described in Eq. (1):$${text{F}}, = ,{text{A}}*{text{dC}}/{text{dt}}$$
(1)
where F is flux of N2O (g ha−1 day−1), A is the area of the chamber, and dC/dt is the change of concentration in time calculated using a linear method of integration by Venterea et al.49.Ammonia volatilization measurementAmmonia volatilization was measured using the open chamber technique, as described by Araújo et al.51. The ammonia chamber was made of a 2-L volume polyethylene terephthalate (PET) bottle. The bottom of the bottle was removed and used as a cap above the top opening to keep the environment controlled, free of insects and other sources of contamination. An iron wire was used to support the plastic jar. A strip of polyfoam (250 mm in length, 25 mm wide, and 3 mm thick) was soaked in 20 ml of acid solution (H2SO4 1 mol dm−3 + glycerine 2% v/v) and fastened to the top, with the bottom end of the foam remaining inside the plastic jar. Inside each chamber there was a 250-mm long wire designed with a hook to support it from the top of the bottle, and wire basket at the bottom end to support a plastic jar (25 mL) that contained the acid solution to keep the foam strip moist during sampling periods (Fig. 7). The ammonia chambers were placed installed in the bucket located in the middle of each experimental block after the last gas sampling of the day and removed before the start of the next gas sampling.Figure 7Mobile ammonia chamber details for ammonia measurement in dung beetle trial. Adapted from Araújo et al.51.Full size imageNutrient cyclingPhotographs of the soil and dung portion of each bucket were taken twenty-four hours after the last day of gas flux measurement sampling to determine the dung removal from single beetle species and their combination. In the section on statistical analysis, the programming and statistical procedures are described. After this procedure, seeds of pearl millet were planted in each bucket. After 5 days of seed germination plants were thinned, maintaining four plants per bucket. Additionally, plants were clipped twice in a five-week interval, with the first cut occurring on October 23rd and the second cut occurring on November 24th, in both experimental years. Before each harvest, plant height was measured twice in the last week. In the harvest day all plants were clipped 10 cm above the ground level. Samples were dried at 55 °C in a forced-air oven until constant weight and ball-milled using a Mixer Mill MM 400 (Retsch, Newton, PA, USA) for 9 min at 25 Hz, and analyzed for total N concentration using a C, H, N, and S analyzer by the Dumas dry combustion method (Vario Micro Cube; Elementar, Hanau, Germany).Statistical analysisTreatments were distributed in a randomized complete block design (RCBD), with three replications. Data were analyzed using the Mixed Procedure from SAS (ver. 9.4., SAS Inst., Cary, NC) and LSMEANS compared using PDIFF adjusted by the t-test (P  More

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.This is a summary of: Zhu, L. et al. Comparable biophysical and biogeochemical feedbacks on warming from tropical moist forest degradation. Nat. Geosci. https://doi.org/10.1038/s41561-023-01137-y (2023). More

Wajnberg, É. et al. (eds) Behavioral Ecology of Insect Parasitoids: From Theoretical Approaches to Field Applications 1st edn. (Blackwell Publishing Ltd, 2008).
Google Scholar 
Godfray, H. C. J. Parasitoids: Behavioral and Evolutionary Ecology (Princeton University Press, 1994).Book 

Google Scholar 
Morris, R. J., Lewis, O. T. & Godfray, H. C. J. Apparent competition and insect community structure: Towards a spatial perspective. Annales Zoologica Fennici 42, 1–14 (2005).
Google Scholar 
Forbes, A. A., Bagley, R. K., Beer, M. A., Hippee, A. C. & Widmayer, H. A. Quantifying the unquantifiable: Why Hymenoptera, not Coleoptera, is the most speciose animal order. BMC Ecol. 18, 1–11 (2018).Article 

Google Scholar 
Hassell, M. P. & Waage, J. K. Host–parasitoid population interactions. Annu. Rev. Entomol. 29, 89–114 (1984).Article 

Google Scholar 
Lafferty, K. D. et al. Parasites in food webs: The ultimate missing links. Ecol. Lett. 11, 533–546 (2008).Article 
PubMed 
PubMed Central 

Google Scholar 
Van Veen, F. J. F., Van Holland, P. D. & Godfray, H. C. J. Stable coexistence in insect communities due to density- and trait-mediated indirect effects. Ecology 86, 3182–3189 (2005).Article 

Google Scholar 
Schmidt, M. H. et al. Relative importance of predators and parasitoids for cereal aphid control. Proc. R. Soc. Lond. Series B Biol. Sci. 270, 1905–1909 (2003).Article 

Google Scholar 
Mills, N. J. & Wajnberg, É. Optimal foraging behavior and efficient biological control methods. In Behavioral Ecology of Insect Parasitoids: From Theoretical Approaches to Field Applications 1st edn (eds Wajnberg, É. et al.) 1–30 (Blackwell Publishing, 2008).
Google Scholar 
Vinson, S. B. Host suitability for insect parasitoids. Annu. Rev. Entomol. 25, 397–419 (1980).Article 

Google Scholar 
Benrey, B. & Denno, R. F. The slow-growth-high-mortality hypothesis: A test using the cabbage butterfly. Ecology 78, 987–999 (1997).
Google Scholar 
Chau, A. & Mackauer, M. Host-instar selection in the aphid parasitoid Monoctonus paulensis (Hymenoptera: Braconidae, Aphidiinae): Assessing costs and benefits. Can. Entomol. 133, 549–564 (2001).Article 

Google Scholar 
Strand, M. R. & Obrycki, J. J. Host specificity of insect parasitoids and predators. Bioscience 46, 422–429 (1996).Article 

Google Scholar 
Vinson, S. B. Host selection by insect parasitoids. Annu. Rev. Entomol. 21, 109–133 (1976).Article 

Google Scholar 
Wang, X. G. & Messing, R. H. Fitness consequences of body-size-dependent host species selection in a generalist ectoparasitoid. Behav. Ecol. Sociobiol. 56, 513–522 (2004).Article 

Google Scholar 
Liu, Z., Xu, B., Li, L. & Sun, J. Host-size mediated trade-off in a parasitoid Sclerodermus harmandi. PLoS ONE 6, e23260 (2011).Article 
ADS 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Wang, X. Y., Yang, Z. Q., Wu, H. & Gould, J. R. Effects of host size on the sex ratio, clutch size, and size of adult Spathius agrili, an ectoparasitoid of emerald ash borer. Biol. Control 44, 7–12 (2008).Article 

Google Scholar 
Hardy, I. C. W., Griffiths, N. T. & Godfray, H. C. J. Clutch size in a parasitoid wasp: A manipulation experiment. J. Anim. Ecol. 61, 121–129 (1992).Article 

Google Scholar 
Scriber, J. M. & Slansky, F. The nutritional ecology of immature insects. Annu. Rev. Entomol. 26, 183–211 (1981).Article 

Google Scholar 
Moreau, J., Benrey, B. & Thiery, D. Assessing larval food quality for phytophagous insects: Are the facts as simple as they appear?. Funct. Ecol. 20, 592–600 (2006).Article 

Google Scholar 
Davidowitz, G., D’Amico, L. J. & Nijhout, H. F. The effects of environmental variation on a mechanism that controls insect body size. Evolut. Ecol. Res. 6, 49–62 (2004).
Google Scholar 
Williams, I. S. Slow-growth, high-mortality-a general hypothesis, or is it?. Ecol. Entomol. 24, 490–495 (1999).Article 

Google Scholar 
Chen, K. & Chen, Y. Slow-growth high-mortality: A meta-analysis for insects. Insect Sci. 25, 337–351 (2018).Article 
PubMed 

Google Scholar 
Waldbauer, G. P. The consumption and utilization of food by insects. Adv. Insect Physiol. 5, 229–288 (1968).Article 

Google Scholar 
Hochuli, D. F. Insect herbivory and ontogeny: How do growth and development influence feeding behaviour, morphology and host use?. Austral. Ecol. 26, 563–570 (2001).Article 

Google Scholar 
Holmes, L. A., Nelson, W. A. & Lougheed, S. C. Food quality effects on instar-specific life histories of a holometabolous insect. Ecol. Evol. 10, 626–637 (2020).Article 
PubMed 
PubMed Central 

Google Scholar 
Kagata, H. & Ohgushi, T. Bottom-up trophic cascades and material transfer in terrestrial food webs. Ecol. Res. 21, 26–34 (2006).Article 

Google Scholar 
Scherber, C. et al. Bottom-up effects of plant diversity on multitrophic interactions in a biodiversity experiment. Nature 468, 553–556 (2010).Article 
ADS 
CAS 
PubMed 

Google Scholar 
Vidal, M. C. & Murphy, S. M. Bottom-up vs top-down effects on terrestrial insect herbivores: A meta-analysis. Ecol. Lett. 21, 138–150 (2018).Article 
PubMed 

Google Scholar 
Harvey, J. A. Factors affecting the evolution of development strategies in parasitoid wasps: The importance of functional constraints and incorporating complexity. Entomol. Exp. Appl. 117, 1–13 (2005).Article 

Google Scholar 
Charnov, E. L., Los-den Hartogh, R. L., Jones, W. T. & van den Assem, J. Sex ratio evolution in a variable environment. Nature 289, 27–33 (1981).Article 
ADS 
CAS 
PubMed 

Google Scholar 
Larson, A. O. The bean weevil and the southern Cowpea weevil in California. United States Department of Agriculture. Technical Bulletin No. 593, Washington, D. C. (1938).Askew, R. R. & Shaw, M. R. Parasitoid communities: their size, structure and development in Insect Parasitoids: 13th Symposium of Royal Entomological Society of London (eds. Waage, J.K. & Greathead, D.J. 225–264 (1986).Holmes, L. A., Nelson, W. A., Dyck, M. & Lougheed, S. C. Enhancing the usefulness of artificial seeds in seed beetle model systems research. Methods Ecol. Evol. 11, 1701–1706 (2020).Article 

Google Scholar 
Ellers, J., Van Alphen, J. J. M. & Sevenster, J. G. A field study of size-fitness relationships in the parasitoid Asobara tabida. J. Anim. Ecol. 67, 318–324 (1998).Article 

Google Scholar 
Wood, S. N. Stable and efficient multiple smoothing parameter estimation for generalized additive models. J. Am. Stat. 99, 673–686 (2004).Article 
MathSciNet 
MATH 

Google Scholar 
Wood, S. N. Generalized Additive Models: An Introduction with R 2nd edn. (Chapman and Hall/CRC, 2017).Book 
MATH 

Google Scholar 
Wood, S. N. Thin-plate regression splines. J. Roy. Stat. Soc. B 65, 95–114 (2003).Article 
MathSciNet 
MATH 

Google Scholar 
R Core Team. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/ (2020). Accessed 3 April 2020.Burnham, K. P. & Anderson, D. R. Model Selection and Inference: A Practical Information-Theoretical Approach 2nd edn. (Springer-Verlag, 2002).MATH 

Google Scholar 
Wood, S. N., Pya, N. & Saefken, B. Smoothing parameter and model selection for general smooth models (with discussion). J. Am. Stat. Assoc. 111, 1548–1575 (2016).Article 
CAS 

Google Scholar 
Bolker, B., & R Development Core Team Tools for general maximum likelihood estimation. Version 1.0.20. (2017). Accessed 4 April 2020.Hothorn, T., Bretz, F. & Westfall, P. Simultaneous inference in general parametric models. Biometical. J. 50, 346–363 (2008).Article 
MathSciNet 
MATH 

Google Scholar 
Rose, N. L., Yang, H., Turner, S. D. & Simpson, G. L. An assessment of the mechanisms for the transfer of lead and mercury from atmospherically contaminated organic soils to lake sediments with particular reference to Scotland, UK. Geochim. Cosmochim. Acta 82, 113–135 (2012).Article 
ADS 
CAS 

Google Scholar 
Holmes, L. A., Nelson, W. A. & Lougheed, S. C. Data from: Food quality effects on instar-specific life histories of a holometabolous insect. Dryad Digital Repository. https://doi.org/10.5061/dryad.d7wm37px7 (2020).Therneau, T. A Package for Survival Analysis in R. R package version 3.2-13. https://CRAN.R-project.org/package=survival. (2021). Accessed 3 April 2020.Efron, B. The Jackknife, the Bootstrap, and Other Resampling Plans (Society for Industrial and Applied Mathematics, 1982).Book 
MATH 

Google Scholar 
Awmack, C. S. & Leather, S. R. Host plant quality and fecundity in herbivorous insects. Annu. Rev. Entomol. 47, 817–844 (2002).Article 
CAS 
PubMed 

Google Scholar 
Clancy, K. M. & Price, P. W. Rapid herbivore growth enhances enemy attack: Sublethal plant defenses remain a paradox. Ecology 68, 733–737 (1987).Article 

Google Scholar 
Loader, C. & Damman, H. Nitrogen content of food plants and vulnerability of Pieris rapae to natural enemies. Ecology 72, 1586–1590 (1991).Article 

Google Scholar 
Uesugi, A. The slow-growth high-mortality hypothesis: Direct experimental support in a leafmining fly. Ecol. Entomol. 40, 221–228 (2015).Article 

Google Scholar 
Feeny, P. Plant apparency and chemical defense. in Biochemical Interaction Between Plants and Insects. 1–40 (Springer, 1976).Teder, T. & Tammaru, T. Cascading effects of variation in plant vigor on the relative performance of insect herbivores and their parasitoids. Ecol. Entomol. 27, 94–104 (2002).Article 

Google Scholar 
Kagata, H., Nakamura, M. & Ohgushi, T. Bottom-up cascade in a tri-trophic system: Different impacts of host-plant regeneration on performance of a willow leaf beetle and its natural enemy. Ecol. Entomol. 30, 58–62 (2005).Article 

Google Scholar 
Vet, L. E. M., Lewis, W. J. & Cardé, R. T. Parasitoid foraging and learning. In Chemical Ecology of Insects 2 (eds Cardé, R. T. & Bell, W. J.) 65–101 (Springer, 1995).Chapter 

Google Scholar 
Ishii, Y. & Shimada, M. Learning predator promotes coexistence of prey species in host-parasitoid systems. Proc. Natl. Acad. Sci. 109, 5116–5120 (2012).Article 
ADS 
CAS 
PubMed 
PubMed Central 

Google Scholar 
Ode, P. J. & Hardy, I. C. Parasitoid sex ratios and biological control. Behavioral ecology of insect parasitoids. In Behavioral Ecology of Insect Parasitoids: From Theoretical Approaches to field applications (eds Wajnberg, E. et al.) 253–291 (Wiley, 2008).Chapter 

Google Scholar 
Xiaoyi, W. & Zhongqi, Y. Behavioral mechanisms of parasitic wasps for searching concealed insect hosts. Acta Ecol. Sin. 28, 1257–1269 (2008).Article 

Google Scholar 
Otten, H., Wäckers, F., Battini, M. & Dorn, S. Efficiency of vibrational sounding in the parasitoid Pimpla turionellae is affected by female size. Anim. Behav. 61, 671–677 (2001).Article 

Google Scholar 
Kaplan, I., Carrillo, J., Garvey, M. & Ode, P. J. Indirect plant-parasitoid interactions mediated by changes in herbivore physiology. Curr. Opin. Insect Sci. 14, 112–119 (2016).Article 
PubMed 

Google Scholar 
Ode, P. J. Plant toxins and parasitoid trophic ecology. Curr. Opin. Insect Sci. 32, 118–123 (2019).Article 
PubMed 

Google Scholar 
Barbosa, P., Gross, P. & Kemper, J. Influence of plant allelochemicals on the tobacco hornworm and its parasitoid, Cotesia congregate. Ecology 72, 1567–1575 (1991).Article 
CAS 

Google Scholar 
Barbosa, P. Natural enemies and herbivore–plant interactions: Influence of plant allelochemicals and host specificity. In Novel Aspects of Insect–Plant Interactions (eds Barbosa, P. & Letourneau, L. D. K.) 201–230 (Wiley, 1988).
Google Scholar 
Ode, P. J. Plant chemistry and natural enemy fitness: Effects on herbivore and natural enemy interactions. Annu. Rev. Entomol. 51, 163–185 (2006).Article 
CAS 
PubMed 

Google Scholar  More

Gillooly, J. F., Brown, J. H., West, G. B., Savage, V. M. & Charnov, E. L. Effects of size and temperature on metabolic rate. Science 293, 2248–2251 (2001).Article 
CAS 

Google Scholar 
Seebacher, F., White, C. R. & Franklin, C. E. Physiological plasticity increases resilience of ectothermic animals to climate change. Nat. Clim. Change 5, 61–66 (2015).Article 

Google Scholar 
Havird, J. C. et al. Distinguishing between active plasticity due to thermal acclimation and passive plasticity due to Q10 effects: why methodology matters. Funct. Ecol. 34, 1015–1028 (2020).Article 

Google Scholar 
Dillon, M. E., Wang, G. & Huey, R. B. Global metabolic impacts of recent climate warming. Nature 467, 704–706 (2010).Article 
CAS 

Google Scholar 
White, C. R., Alton, L. A., Bywater, C. L., Lombardi, E. J. & Marshall, D. J. Metabolic scaling is the product of life history optimization. Science 377, 834–839 (2022).Article 
CAS 

Google Scholar 
Savage, V. M., Gilloly, J. F., Brown, J. H. & Charnov, E. L. Effects of body size and temperature on population growth. Am. Nat. 163, 429–441 (2004).Article 

Google Scholar 
Bernhardt, J. R., Sunday, J. M. & O’Connor, M. I. Metabolic theory and the temperature–size rule explain the temperature dependence of population carrying capacity. Am. Nat. 192, 687–697 (2018).Article 

Google Scholar 
Damuth, J. Population density and body size in mammals. Nature 290, 699–700 (1981).Article 

Google Scholar 
Schuster, L., Cameron, H., White, C. R. & Marshall, D. J. Metabolism drives demography in an experimental field test. Proc. Natl Acad. Sci. USA 118, e2104942118 (2021).Article 
CAS 

Google Scholar 
Amarasekare, P. & Coutinho, R. M. The intrinsic growth rate as a predictor of population viability under climate warming. J. Anim. Ecol. 82, 1240–1253 (2013).Article 

Google Scholar 
Amarasekare, P. & Savage, V. A framework for elucidating the temperature dependence of fitness. Am. Nat. 179, 178–191 (2012).Article 

Google Scholar 
Lande, R. Risks of population extinction from demographic and environmental stochasticity and random catastrophes. Am. Nat. 142, 911–927 (1993).Article 

Google Scholar 
Comeault, A. A. & Matute, D. R. Temperature-dependent competitive outcomes between the fruit flies Drosophila santomea and Drosophila yakuba. Am. Nat. 197, 312–323 (2021).Article 

Google Scholar 
Davis, A. J., Jenkinson, L. S., Lawton, J. H., Shorrocks, B. & Wood, S. Making mistakes when predicting shifts in species range in response to global warming. Nature 391, 783–786 (1998).Article 
CAS 

Google Scholar 
Davis, A. J., Lawton, J. H., Shorrocks, B. & Jenkinson, L. S. Individualistic species responses invalidate simple physiological models of community dynamics under global environmental change. J. Anim. Ecol. 67, 600–612 (1998).Article 

Google Scholar 
Gilman, S. E., Urban, M. C., Tewksbury, J., Gilchrist, G. W. & Holt, R. D. A framework for community interactions under climate change. Trends Ecol. Evol. 25, 325–331 (2010).Article 

Google Scholar 
Janča, M. & Gvoždík, L. Costly neighbours: heterospecific competitive interactions increase metabolic rates in dominant species. Sci. Rep. 7, 5177 (2017).Article 

Google Scholar 
Pettersen, A. K., Hall, M. D., White, C. R. & Marshall, D. J. Metabolic rate, context-dependent selection, and the competition–colonization trade-off. Evol. Lett. 4, 333–344 (2020).Article 

Google Scholar 
DeLong, J. P., Hanley, T. C. & Vasseur, D. A. Competition and the density dependence of metabolic rates. J. Anim. Ecol. 83, 51–58 (2014).Article 

Google Scholar 
Reid, D., Armstrong, J. D. & Metcalfe, N. B. Estimated standard metabolic rate interacts with territory quality and density to determine the growth rates of juvenile Atlantic salmon. Funct. Ecol. 25, 1360–1367 (2011).Article 

Google Scholar 
Ayala, F. J. in Essays in Evolution and Genetics in Honor of Theodosius Dobzhansky (eds Hecht, M. K. & Steere, W. C.) 121–158 (Springer, 1970).Atkinson, W. D. & Shorrocks, B. Aggregation of larval Diptera over discrete and ephemeral breeding sites: the implications for coexistence. Am. Nat. 124, 336–351 (1984).Article 

Google Scholar 
McKenzie, J. A. & McKechnie, S. W. A comparative study of resource utilization in natural populations of Drosophila melanogaster and D. simulans. Oecologia 40, 299–309 (1979).Article 
CAS 

Google Scholar 
Alton, L. A. et al. Developmental nutrition modulates metabolic responses to projected climate change. Funct. Ecol. 34, 2488–2502 (2020).Article 

Google Scholar 
Mitchell, K. A. & Hoffmann, A. A. Thermal ramping rate influences evolutionary potential and species differences for upper thermal limits in Drosophila. Funct. Ecol. 24, 694–700 (2010).Article 

Google Scholar 
Overgaard, J., Kristensen, T. N., Mitchell, K. A. & Hoffmann, A. A. Thermal tolerance in widespread and tropical Drosophila species: does phenotypic plasticity increase with latitude? Am. Nat. 178, S80–S96 (2011).Article 

Google Scholar 
Kellermann, V. et al. Comparing thermal performance curves across traits: how consistent are they? J. Exp. Biol. 222, jeb193433 (2019).Article 

Google Scholar 
Terblanche, J. S., Clusella-Trullas, S. & Chown, S. L. Phenotypic plasticity of gas exchange pattern and water loss in Scarabaeus spretus (Coleoptera: Scarabaeidae): deconstructing the basis for metabolic rate variation. J. Exp. Biol. 213, 2940–2949 (2010).Article 

Google Scholar 
Tewksbury, J. J., Huey, R. B. & Deutsch, C. A. Putting the heat on tropical animals. Science 320, 1296–1297 (2008).Article 
CAS 

Google Scholar 
Bos, M., Burnet, B., Farrow, R. & Woods, R. A. Mutual facilitation between larvae of the sibling species Drosophila melanogaster and D. simulans. Evolution 31, 824–828 (1977).Article 
CAS 

Google Scholar 
Arthur, W. On the complexity of a simple environment: competition, resource partitioning and facilitation in a two-species Drosophila system. Phil. Trans. R. Soc. B 313, 471–508 (1986).
Google Scholar 
Hodge, S., Mitchell, P. & Arthur, W. Factors affecting the occurrence of facilitative effects in interspecific interactions: an experiment using two species of Drosophila and Aspergillus niger. Oikos 87, 166–174 (1999).Article 

Google Scholar 
Bath, E., Morimoto, J. & Wigby, S. The developmental environment modulates mating-induced aggression and fighting success in adult female Drosophila. Funct. Ecol. 32, 2542–2552 (2018).Article 

Google Scholar 
Thibert, J., Farine, J. P., Cortot, J. & Ferveur, J. F. Drosophila food-associated pheromones: effect of experience, genotype and antibiotics on larval behavior. PLoS ONE 11, e0151451 (2016).Article 

Google Scholar 
Chown, S. L. et al. Scaling of insect metabolic rate is inconsistent with the nutrient supply network model. Funct. Ecol. 21, 282–290 (2007).Article 

Google Scholar 
Becker, R. A., Wilks, A. R. & Brownrigg, R. mapdata: extra map databases. R version 2.3.0 https://CRAN.R-project.org/package=mapdata (2018).R Core Team R: A Language and Environment for Statistical Computing (R Foundation for Statistical Computing, 2022).Brooks, M. E. et al. glmmTMB balances speed and flexibility among packages for zero-inflated generalized linear mixed modeling. R J. 9, 378–400 (2017).Article 

Google Scholar 
Bates, D., Mächler, M., Bolker, B. & Walker, S. Fitting linear mixed-effects models using lme4. J. Stat. Softw. 67, 1–48 (2015).Article 

Google Scholar 
Bolker, B. & R Development Core Team bbmle: tools for general maximum likelihood estimation. R version 1.0.25 https://CRAN.R-project.org/package=bbmle (2022).Kuznetsova, A., Brockhoff, P. B. & Christensen, R. H. B. lmerTest package: tests in linear mixed effects models. J. Stat. Softw. 82, 1–26 (2017).Article 

Google Scholar 
Fox, J. & Weisberg, S. An R Companion to Applied Regression 3rd edn (Sage, 2019).Hartig, F. DHARMa: residual diagnostics for hierarchical (multi-level/mixed) regression models. R version 0.4.6 https://CRAN.R-project.org/package=DHARMa (2022).Messamah, B., Kellermann, V., Malte, H., Loeschcke, V. & Overgaard, J. Metabolic cold adaptation contributes little to the interspecific variation in metabolic rates of 65 species of Drosophilidae. J. Insect Physiol. 98, 309–316 (2017).Article 
CAS 

Google Scholar 
Chamberlain, S. et al. rgbif: interface to the global biodiversity information facility API. R version 3.7.3 https://CRAN.R-project.org/package=rgbif (2022).Fick, S. E. & Hijmans, R. J. WorldClim 2: new 1‐km spatial resolution climate surfaces for global land areas. Int. J. Climatol. 37, 4302–4315 (2017).Article 

Google Scholar 
Hijmans, R. J. raster: geographic data analysis and modeling. R version 3.6-3 https://CRAN.R-project.org/package=raster (2022).Alton, L. A. & Kellermann, V. Data for “Interspecific interactions alter the metabolic costs of climate warming”. Zenodo https://doi.org/10.5281/zenodo.7475922 (2023).White, C. R. et al. Geographical bias in physiological data limits predictions of global change impacts. Funct. Ecol. 35, 1572–1578 (2021).Article 

Google Scholar  More