Ecology

Glass eels collection (Anguilla anguilla, L.)Three hundred European glass eels ( More

External datasetsWoody biomass carbon dataThe dataset by ref. 16 maps annual global woody biomass carbon densities for 2000–2019 at a spatial resolution of ~10 km. The annual estimates represent averages for the tropical regions and growing-season (April–October) averages for the extra-tropical regions. Ref. 16 analyse global trends of gains and losses in woody biomass carbon for 2000–2019. Overall, they find that grid cells with (significant) net gains of vegetation carbon are by a factor of 1.4 more abundant than grid cells with net losses of vegetation carbon, indicating that there is a global greening trend when only considering the areal extent of biomass gains and not the magnitude of carbon gains. Their regionally distinct analysis of trends shows that almost all regions, except for the tropical moist forests in South America and parts of Southeast Asia, experienced net gains in biomass carbon. On the country scale, the largest net increase in biomass carbon is shown in China, which is mainly attributed to the large-scale afforestation programs in the southern part of the country and increased carbon uptake of established forests. On the other hand, the largest vegetation carbon losses are shown for Brazil and Indonesia, which is partly attributed to deforestation, degradation, and drought events. All of the mentioned trends have been found to be significant16. The decreasing carbon sink in Brazil is in line with ref. 44, who, considering both natural and anthropogenic fluxes, show that the southeastern Amazon has even turned from a carbon sink to a carbon source, mainly owing to fire emissions from forest clearing. Isolating carbon fluxes in intact, old-growth Amazonian rainforests (i.e., SLAND,B), ref. 45 also find evidence for a significantly decreasing carbon sink due to the negative effects of increasing temperatures and droughts on carbon uptake since the 1990s.The dataset was remapped to the BLUE resolution of 0.25∘ through conservative remapping (i.e., area-weighted averaging).ERA-5 dataThe ERA-5 variables were downloaded from the Copernicus Climate Data Store (https://cds.climate.copernicus.eu/cdsapp#!/home). Monthly air temperature (Ta) at 2 m height was averaged over each year, and annual precipitation was calculated by taking the sum of the monthly total precipitation (P). Both variables were regridded from the original resolution of ~0.1° to 0.25° resp. to the TRENDY resolution of 0.5° through conservative remapping.TRENDY dataWe used the TRENDY model ensemble version 8 (conducted for the 2019 GCB; ref. 8). We used net biome production (NBP) and annual vegetation carbon stocks (cVeg) for 2000–2018 from four different model setups (S2, S3, S5, and S6) and eight resp. 13 DGVMs (depending on the data available). The selection of DGVMs is done as in ref. 19 (Supplementary Tab. 3), but we included one additional model (ISAM) for the S2 simulations. The terrestrial biomass carbon sink (SLAND,B) was calculated for 13 DGVMs following the GCB 2020 approach, i.e., from the S2 simulation, which is the simulation without LULCC (i.e., fixed pre-industrial land cover) under transient environmental conditions (climate, nitrogen deposition, CO2 evolution). SLAND,B is the annual difference of cVeg and makes no statements about the further fate of biomass if cVeg decreases. SLAND,B, therefore, should not be interpreted as equivalent to the flux to/from the atmosphere, since parts of cVeg may be transferred to litter, dead wood, or soil. The same applies to our BLUE estimates of SLAND,B, ensuring comparability between our BLUE estimates and the TRENDY estimates. Increases (decreases) of cVeg between two years are a net uptake (release) of carbon from the terrestrial biosphere. The global sums of biomass carbon stocks under transient climate and CO2 were calculated from the S3 setup (LULCC under historical environmental conditions), whereas the S5 setup provides biomass carbon under constant present-day environmental forcing (closest to the classical bookkeeping approach). In line with the GCB, ELUC was calculated under historical environmental conditions as the difference in NBP between the S2 and S3 simulations (ELUC = NBP_S2 – NBP_S3). ELUC under constant present-day environmental forcing was calculated as the difference in NBP between the S6 (fixed pre-industrial land cover under present-day environmental forcing) and S5 simulations (ELUC = NBP_S6 – NBP_S5)19. All datasets were remapped to a common resolution of 0.5∘ through conservative remapping (area-weighted average) for the data analysis.Assimilation of observed woody biomass carbon in BLUEThe observed woody biomass carbon densities by ref. 16 are assimilated in BLUE in several steps.Carbon transfer in the default setup of BLUEThe BLUE simulation is started in AD 850. Biomass and soil vegetation carbon densities are based on ref. 17, which are converted to exponential time constants. A detailed explanation of the exponential model can be found in ref. 5.While in the default setup, changes are only due to LULCC, our assimilation approach now introduces environmental effects on woody vegetation carbon by assimilating the observed woody biomass carbon densities in BLUE from 2000 onward according to the methodological considerations explained below.Calculation of woody biomass carbon densities for different land cover types and PFTsWithin each 0.25° cell of the global grid, the (remapped) woody biomass carbon density from ref. 16 must be the sum of woody biomass carbon stored in all woody PFTs of all woody land cover types. The distribution of the woody biomass carbon across PFTs and land cover types is achieved by distributing the observed (i.e., actual) woody biomass carbon densities (ρBa) from ref. 16 across the two land cover types (j) and the eight PFTs (l) that can be woody vegetation (primary land, called virgin, “v” in BLUE and secondary, “s”, land) according to the fraction of total woody biomass carbon (fB) contained in each land cover type and each PFT (fB,j,l) as estimated by BLUE. fB,j,l varies for different PFTs and land cover types, depending on their history of LULCC and their potential for carbon uptake (i.e., the potential carbon densities).fB,j,l is extracted from the default simulations for the first year of the time series (i.e., 2000) and calculated for subsequent years from the BLUE simulations using the assimilated woody vegetation carbon densities for that year:$${f}_{B,j,l}(t)=frac{{C}_{B,j,l}(t)}{{C}_{B}(t)}$$
(1)
where CB is the woody biomass carbon stock.Consequently, the assimilated woody biomass carbon stock per cover type and PFT (CB_as,j,l) at each time step can be calculated as:$${C}_{B_as,j,l}(t)={rho }_{Ba}(t);*;A;*;{f}_{B,j,l}(t)$$
(2)
with j{v, s}; l{1. . 8}; t{2000. . 2019}. A is the area per grid cell.Thresholds for excluding inconsistent woody biomass carbon densitiesWe eliminate unrealistically large values for woody biomass carbon densities that our assimilation framework produces. Woody biomass carbon densities in BLUE that exceed the highest value (~374 t ha−1) of the original dataset indicate inconsistencies between the observed woody biomass carbon estimates and the fractional grid cell areas per PFT and land cover types that BLUE simulates. To account for uncertainties related to the criteria for exclusion of grid cells, multiple threshold approaches are applied and the results are compared. To maintain a temporally and spatially consistent time series of woody biomass carbon, grid cells that are excluded according to the chosen threshold approach are interpolated through linear barycentric interpolation. A first approach relies on a uniform upper threshold of More

We focused on the Northern High Latitudes (NHL, latitude > 50°N, excluding Greenland) due to their importance for carbon (CO2-C, the same hereafter)-climate feedbacks in the Earth system. To minimize the potential human influence on the CO2 cycle, we excluded areas under agricultural management (croplands, cropland/natural vegetation mosaic, and urban types), and considered only pixels of natural vegetation defined from the MODIS MCD12Q1 (v006) based IGBP land cover classification. Our main focus was the NHL permafrost region because permafrost plays a critical role in the ecology, environment, and society in the NHL. Permafrost, or permanently frozen ground, is defined as ground (soil, sediment, or rock) that remains at or below 0 °C for at least two consecutive years. The occurrence of permafrost is primarily controlled by temperature and has a strong effect on hydrology, soils, and vegetation composition and structure. Based on the categorical permafrost map from the International Permafrost Association58, the permafrost region (excluding permanent snow/ice and barren land), including sporadic (10–50%), discontinuous (50–90%), and continuous ( >90%) permafrost, encompasses about 15.7 × 106 km2, accounts for 57% of the NHL study dominion, and is dominated by tundra (shrubland and grass) and deciduous needleleaf (i.e., larch) forest that is regionally abundant in Siberia. The NHL non-permafrost region covers about 11.9 × 106 km2 and is dominated by mixed and evergreen needleleaf boreal forests (Fig. S1).Atmospheric CO2 inversions (ACIs)ACIs provide regionally-integrated estimates of surface-to-atmosphere net ecosystem CO2 exchange (NEEACI) fluxes by utilizing atmospheric CO2 concentration measurements and atmospheric transport models59. ACIs differ from each other mainly in their underlying atmospheric observations, transport models, spatial and temporal flux resolutions, land surface models used to predict prior fluxes, observation uncertainty and prior error assignment, and inversion methods. We used an ensemble mean of six different ACI products, each providing monthly gridded NEEACI at 1-degree spatial resolution, including Carbon‐Tracker 2019B (2000-2019, CT2019)60, Carbon‐Tracker Europe 2020 (2000–2019, CTE2020)61, Copernicus Atmosphere Monitoring Service (1979–2019, CAMS)62, Jena CarboScope (versions s76_v4.2 1976–2017, and s85_v4.2 1985-2017)63,64, and JAMSTEC (1996–2017)65. The monthly gridded ensemble mean NEEACI at 1-degree spatial resolution was calculated using the available ACIs from 1980-2017. Monthly ACI ensemble mean NEEACI data were summed to seasonal and annual values, and used to calculate the spatial and temporal trends of net CO2 uptake, and to investigate its relationship to climate and environmental controls.Productivity datasetDirect observations of vegetation productivity do not exist at a circumpolar scale. We therefore used two long-term gridded satellite-based estimates of vegetation productivity, including gross primary production (GPP) derived using a light use efficiency (LUE) approach (LUE GPP, 1982–1985)21,66 and satellite observations of Normalized Difference Vegetation Index (NDVI) from the Global Inventory Modeling and Mapping Studies (GIMMS NDVI, 1982–1985)67. LUE GPP (monthly, 0.5° spatial resolution, 1982–2015) is calculated from satellite observations of NDVI from the Advanced Very High-Resolution Radiometer (AVHRR; 1982 to 2015) combined with meteorological data, using the MOD17 LUE approach. LUE GPP has been extensively validated with a global array of eddy-flux tower sites68,69,70 and tends to provide better estimates in ecosystems with greater seasonal variability at high latitudes. Following66,71, we used the ensemble mean of GPP estimates from three of the most commonly used meteorological data sets: National Centers for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) reanalysis; NASA Global Modeling and Assimilation Office (GMAO) Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2); and European Center for Medium-Range Weather Forecasting (ECMWF). GIMMS NDVI (bimonthly, 1/12 spatial resolution, 1982–2015) provides the longest satellite observations of vegetation “greenness”, and is widely used in studies of phenology, productivity, biomass, and disturbance monitoring as it has proven to be an effective surrogate of vegetation photosynthetic activity72.The gridded GPP data were resampled to 1-degree resolution at monthly time scales, to be consistent with NEEACI, and used to test (H1) whether greater temperature sensitivity of vegetation productivity explains the different trends in net CO2 uptake across the NHL. LUE GPP was also used to calculate monthly total ecosystem respiration (TER) as the difference between GPP and NEEACI (i.e., TERresidual =  GPP– NEEACI) from 1982-2015, as global observations of respiration do not exist. The NEEACI, GPP and TERresidual were used as observation-constrained top-down CO2 fluxes to investigate mechanisms underlying the seasonal CO2 dynamics in the structural equation modeling and additional decision tree-based analysis.Eddy Covariance (EC) measurements of bottom-up CO2 fluxesA total of 48 sites with at least three years of data representing the major NHL ecosystems were obtained from the FLUXNET2015 database (Table S1 and Fig. S1). EC measurements provide direct observations of net ecosystem CO2 exchange (NEE) and estimate the GPP and TER flux components of NEE using other climate variables. Daily GPP and TER were estimated as the mean value from both the nighttime partitioning method73 and the light response curve method74. More details on the flux partitioning and gap-filling methods used are provided by75. Daily fluxes were summed into seasonal and annual values and used to compare with trends from ACIs (Fig. S7), to estimate the climate and environmental controls on the CO2 cycle in the pathway analysis (Fig. 5), and to calculate the net CO2 uptake sensitivity to spring temperature (Fig. S14).Ensemble of dynamic global vegetation models (TRENDY simulations)The TRENDY intercomparison project compiles simulations from state-of-the-art dynamic global vegetation models (DGVMs) to evaluate terrestrial energy, water, and net CO2 exchanges76. The DGVMs provide a bottom-up approach to evaluate terrestrial CO2 fluxes (e.g., net biome production [NBP]) and allow deeper insight into the mechanisms driving changes in carbon stocks and fluxes. We used monthly NBP, GPP, and TER (autotrophic + heterotrophic respiration; Ra + Rh) from ten TRENDY v7 DGVMs76, including CABLE-POP, CLM5.0, OCN, ORCHIDEE, ORCHIDEE-CNP, VISIT, DLEM, LPJ, LPJ-GUESS, and LPX. We analyzed the “S3” simulations that include time-varying atmospheric CO2 concentrations, climate, and land use. All simulations were based on climate forcing from the CRU-NCEPv4 climate variables at 6-hour resolution. CO2 flux outputs were summarized monthly at 1-degree spatial resolution from 1980 to 2017. Monthly ensemble mean NBP, GPP, and TER were summed to seasonal and annual values, and then used to compare with observation-constrained ACI top-down CO2 fluxes (Figs. 4 and 5).Satellite data-driven carbon flux estimates (SMAP L4C)We also used a much finer spatio-temporal simulation of carbon fluxes from the NASA Soil Moisture Active Passive (SMAP) mission Level 4 Carbon product (L4C) to quantify the temperature and moisture sensitivity of NHL CO2 exchange77. The SMAP L4C provides global operational daily estimates of NEE and component CO2 fluxes for GPP and TER at 9 km resolution since 2015; whereas, an offline version of the L4C model provides a similar Nature Run (NR) carbon flux record over a longer period (2000-present), but without the influence of SMAP observational inputs. The L4C model has been calibrated against FLUXNET tower CO2 flux measurements and shows favorable performance and accuracy in high latitude regions4,77. In this analysis, daily gridded CO2 fluxes at 9-km resolution from the L4C NR record were summed to seasonal and annual values, and used to calculate the sensitivity of net C uptake in response to spring temperature (Fig. S14).CO2 fluxes in this analysis are defined with respect to the biosphere so that a positive value indicates the biosphere is a net sink of CO2 absorbed from the atmosphere. The different data products described above use different terminology (e.g., NEE, NBP) with slightly different meanings; however, they all provide estimates of net land-atmosphere CO2 exchange78.Climate, tree cover, permafrost, and soil moisture dataMonthly gridded air temperatures at 0.5-degree spatial resolution from 1980 to 2017 were obtained from the Climate Research Unit (CRU TS v4.02) at the University of East Anglia79. Air temperature was summarized at seasonal and annual scales to calculate temperature sensitivities of net CO2 uptake and to investigate the mechanism underlying the seasonal CO2 dynamics.Percent tree cover (%TC) at 0.05-degree spatial resolution was averaged over a 35-year (1982-2016) period using annual %TC layers derived from the Advanced Very High-Resolution Radiometer (AVHRR) (Fig. 1a)42. %TC was binned using 5% TC intervals to assess its relation to net CO2 uptake, or aggregated at a regional scale (e.g., TC  > 50% or TC  90%), discontinuous permafrost (DisconP, 10% < P  90%), discontinuous (DisconP, 10% < P  0.05 indicate a good fitting model), Bentler’s comparative fit index (CFI, where CFI ≈ 1 indicates a good fitting model), and the root mean square error of approximation (RMSEA; where RMSEA ≤ 0.05 and p  > 0.1 indicate a good fitting model). The standardized regression coefficient can be interpreted as the relative influences of exogenous (independent) variables. The R2 indicates the total variation in an endogenous (dependent) variable explained by all exogenous (independent) variables.Direct and legacy effects of temperature on seasonal net CO2 uptakeBecause landscape thawing and snow conditions regulate the onset of vegetation growth and influence the seasonal and annual CO2 cycles in the NHL24,84, we also analyzed the legacy effects of spring (May–Jun) temperature on seasonal net CO2 uptake. We regressed seasonal and annual net CO2 uptake from the site-level EC observations, regional-level ACI ensemble, and the TRENDY NBP ensemble against spring (May-June) air temperature. For EC observations, net CO2 uptake (i.e., NEE) and air temperature were summarized from site-level measurements. For the ACIs and TRENDY ensemble, net CO2 uptake (i.e., NEEACI and NBP) was summarized as regional means from the ACIs and TRENDY ensemble outputs, and air temperature was summarized as regional means from CRU temperature. The slope of the regression line was interpreted as the spring temperature sensitivity of the CO2 cycle. Simple linear regression was used here mainly due to the strong influence of spring temperature on the seasonal and annual CO2 cycle in NHL ecosystems30. Temperature sensitivity (γ: g C m−2 day−1 K−1) is the change in net CO2 flux (g C m−2 day−1) in response to a 1-degree temperature change. The sensitivity of net CO2 uptake to warm spring anomalies was calculated for different seasons (EGS, LGS, and annual) and regions (i.e., permafrost and non-permafrost), and the T-test was used to test for the difference in γ among different regions, seasons, and datasets. Similarly, direct effects of temperature on net CO2 uptake were calculated using the same season data (Fig. S14).Observationally-constrained estimates (EC and ACIs) showed that the sensitivity of net CO2 uptake in the EGS to spring temperature is positive (γ  > 0) and not statistically different (p  > 0.05) between permafrost and non-permafrost regions (({gamma }_{{ACI}}^{{np}})=0.125 ± 0.020 gC m−2 d−1 K−1; ({gamma }_{{EC}}^{{np}}) = 0.052 ± 0.013 gC m−2 d−1 K−1). In contrast, the sensitivity of net CO2 uptake in LGS to spring temperature is negative (γ  More

Experimental designA 2-year field experiment was conducted at the Modern Agricultural Research and Development Base of Henan Province (113° 35′–114° 15′ E, 34° 53′–35° 11′ N). In order to enhance the diversity of LAI data, a split-plot design with a variety of field management measures and three replications was selected for the experiment (Fig. 1). The size of each experiment plot was 40 m2, the soil texture was predominantly sandy loam and sandy clay loam, as determined by textural analysis of soil samples collected before planting. Maize cultivar Dedan-5 was used in the experiment, which was planted on June 12, 2019, and June 20, 2020, with a row spacing of 42 cm and a planting density of 7 seedlings·m−2. The soil and cultivar in field experiments were representatives of those in the region. The irrigation, pesticide, and herbicide control practices followed local management for maize production.Figure 1The experimental design.Full size imageLAI measurements and UAV-based image acquisitionThe measurements of LAI were conducted at four growth stages including the tasseling stage (TS), flowering stage (FS), grain-filling stage (GS), and milk-ripe stage (MS) of maize in 2019 and 2020, a total of 264 LAI data of maize were collected during the 2-year field trial (Table 1). In order to reduce the impact of plant variability, the random sampling method was used to collect LAI samples. For each plot, three plants were randomly selected to measure the total green leaf area with the non-destructive portable leaf area meter (Laser Area Meter CI-203; CID Inc.). And the average leaf area of selected plants represented the single plant leaf area in each experiment plot. The LAI of each plot wasTable 1 Description of samplings.Full size table$$mathrm{LAI}=mathrm{LA}*mathrm{D}$$
(1)
where (mathrm{LA}) is the leaf area of a single plant in each plot; (mathrm{D}) is the planting density in one square meter.PHANTOM 4 PRO (DJI-Innovations Inc., Shenzhen, China) is a multi-rotor UAV equipped with a 20-megapixel visible-light camera that was employed to capture digital images. Aerial observations were conducted on the same dates as the LAI measurements, which was between 10:30 a.m. and 2:00 p.m. local time when the solar zenith angle was minimal. The UAV was flown automatically based on preset flight parameters and waypoints, with a forward overlap of 80% and a side overlap of 60%. A three-axis gimbal integrated with the inertial navigation system stabilized the camera, the automatic camera mode with fixed ISO (100) and a fixed exposure was used during the flight. Altogether, 4192 images were taken in eight flights from a flight height of 29.36 m above ground, with a spatial resolution of 0.008 m.The measurements of maize LAI were carried out with permission from the Modern Agricultural Research and Development Base of Henan Province. All experiments were carried out in accordance with relevant institutional, national, and international guidelines and legislation.Image pre-processingDJI Terra (version 2.3.3) was used to generate ortho-rectified images based on the structure from motion algorithms and a mosaic blending model. The main procedures are as follows: (1) extract feature points and match features according to the longitude, latitude, elevation, roll angle, pitch angle, and heading angle of each image; (2) build dense 3D point clouds by using dense multi-view stereo matching algorithm; (3) build a 3D polygonal mesh based on the vector relationship between each point in the dense cloud; (4) establish a 3D model with both external image and internal structure by merging the mosaic image into the 3D model; (5) generate digital orthophoto map (DOM).Vegetation indices (VIs) derived from the UAV-based digital imageryDigital imagery records the intensity of visible red (R), green (G), and blue (B) bands in individual pixels24. In order to enhance the vegetation parameters contained in the digital image, fourteen commonly used RGB-based VIs were collected, and their correlation with the LAI of maize at different growth stages was evaluated. Table 2 shows the detailed information of the selected RGB-based VIs.Table 2 RGB-based VIs for LAI estimation.Full size tableCentered on the point where LAI was measured, regions of interests (ROIs) with a size of 100*100 were clipped from the digital image. Python 3.7.3 was used for extracting the R, G, B information of maize and computing the RGB-based VIs from ROIs. In order to reduce the effects of light and shadow, the R, G, B color space of the image was normalized according to the followings:$$mathrm{r}=frac{R}{R+G+B}$$
(2)
$$g=frac{G}{R+G+B}$$
(3)
$$b=frac{B}{R+G+B}$$
(4)
where r, g, and b are the normalized values. R, G, B are the pixel values from the digital images based on each band.Pearson correlation analysisBefore regression analysis, the Pearson correlation analysis was performed to determine the relationship between maize LAI and different RGB-based VIs extracted from the digital image. Pearson correlation coefficient ((mathrm{r})) reflects the degree of linear correlation between two variables, which is between − 1 and 1. The calculation formula of Pearson correlation coefficient was expressed as follows:$$mathrm{r}= frac{sum_{i=1}^{n}left({X}_{i}-overline{X }right)left({Y}_{i}-overline{Y }right)}{sqrt{sum_{i=1}^{n}{left({X}_{i}-overline{X }right)}^{2}}sqrt{sum_{i=1}^{n}{left({Y}_{i}-overline{Y }right)}^{2}}}$$
(5)
where (X), (mathrm{Y}) are variables, (n) is the number of variables.Regression methodsLinear regression (LR)Linear regression is an approach for modelling the relationship between dependent and independent variables. The case of one independent variable is called unary linear regression (ULR), the expressions can be expressed as follows:$$mathrm{y}={beta }_{0}+{beta }_{1}x+varepsilon $$
(6)
where (varepsilon ) is deviation, which satisfies the normal distribution. (x), (mathrm{y}) are variables. ({beta }_{0}), ({beta }_{1}) are the intercept and slope of the regression line, respectively.For more than one independent variable, the regression process is called multiple linear regression (MLR), the expressions can be expressed as:$$mathrm{y}={beta }_{0}+{beta }_{1}{x}_{1}+{beta }_{2}{x}_{2}+dots +{beta }_{n}{x}_{n}$$
(7)
where ({x}_{1}),( {x}_{2}), …, ({x}_{n}), (mathrm{y}) are variables, ({beta }_{0}), ({beta }_{1}), ({beta }_{2}), …, ({beta }_{n}) are coefficients that determined by least square method and gradient descent method38.The RGB-based VIs with the highest Pearson correlation coefficient was used to establish the ULR model, and VIs with a correlation coefficient higher than 0.7 were used to establish the MLR model. In each growth stage, 70% of observation data were randomly selected for establishing models, and the remaining 30% of data were used as the testing dataset to assess the model performance.Back propagation neural networks (BPNN)In this study, a three-layer BPNN model was established for LAI estimation (Fig. 2). RGB-based VIs with a correlation coefficient higher than 0.7 were selected as the input variables. Tan-Sigmoid activation function was used in the hidden layer, and the Levenberg–Marquardt algorithm was selected as the training function. The maximum epoch of BPNN training was set to 1000, the learning rate was set to 0.005, and the MSE was set to 0.001. The observation data set was split into the training set and the testing dataset with a ratio of 7:3. The training dataset was used to fit the weights and bias of the BPNN model, the testing dataset was used to evaluate the model performance. Before training, data normalization was conducted for the input and output variables, and the denormalization was required to convent the output variable back into the original units after training.Figure 2Three-layer BPNN model.Full size imageRandom forest (RF)RF is a non-parametric ensemble ML method that operates by constructing a multitude of decision trees at training time and outputting the average prediction of the individual trees (Fig. 3). The bootstrapping approach was used to collect different sub-training data from the input training dataset to construct individual decision trees.Figure 3Random forest model.Full size imageThe construction process of RF regression model is as follows:

(1)

The value of (mathrm{n}_mathrm{estimators}) was tested from 50 to 1000 in increments of 50, and the value of 500 was finally selected according to higher R2 and lower RMSE.

(2)

At each node per tree, (mathrm{m}_mathrm{try}) RGB-based VIs was randomly selected from all 14 vegetation indices, and the best split was chosen according the lowest Gini Index. (mathrm{m}_mathrm{try}) was tested from 3 to 10, and the final value was 6.

(3)

The other parameters in the RF model were kept as default values according to the (mathrm{RandomForestRegressor}) function in (mathrm{Scikit}-mathrm{learn library}).

(4)

For each tree, the data splitting process in each internal node was repeated from the root node until a pre-defined stop condition was reached.

(5)

Similar with LR and BPNN model, the RGB-based VIs with a correlation coefficient higher than 0.7 were selected as the input variables, and the output variable is LAI.

Data analysis and performance evaluationThe repeated random sampling validation method was used to evaluate the generalization performance of different models. The training and testing dataset were randomly split 500 times. For each split, the LR, BPNN, and RF models were fitted to the training dataset, and the estimation accuracy was evaluated using the testing dataset. The coefficient of determination (R2), root mean square error (RMSE), and Akaike information criterion (AIC) of the training dataset were used for the assessment of models39, and the estimation accuracy was evaluated by R2 and RMSE of the testing dataset. Mathematically, a higher R2 corresponds to a smaller RMSE, and thus represents better model performance. The procedures of LAI inversion using UAV-based digital imagery and ML methods were shown in Fig. 4.Figure 4Flowchart of LAI inversion using UAV-based remote sensing and ML methods.Full size imageThe construction and evaluation of models was performed using Python 3.7.3 in Windows 10 operating system with Intel Core i7-9700 processor, 3.00 GHz CPU, and 32 GB RAM. The processing software is Spyder. The statistical analysis and figure plotting were performed in R × 64 4.0.3. More

Woodward SE, Krekhno Z, Finlay BB. Here, There, and Everywhere: How Pathogenic Escherichia coli Sense and Respond to Gastrointestinal Biogeography. Cell Microbiol. 2019;21:e13107.CAS 
Article 

Google Scholar 
Collins JW, Keeney KM, Crepin VF, Rathinam VAK, Fitzgerald KA, Finlay BB, et al. Citrobacter rodentium: Infection, Inflammation and the Microbiota. Nat Rev Microbiol. 2014;12:612–23.CAS 
Article 

Google Scholar 
Mowat AM, Agace WW. Regional Specialization Within the Intestinal Immune System. Nat Rev Immunol. 2014;14:nri3738.Article 

Google Scholar 
Population Dynamics – Latest Research and News | Nature [Internet]. [cited 2017 Nov 5]. Available from: https://www.nature.com/subjects/population-dynamicsAbel S, Abel zur Wiesch P, Davis BM, Waldor MK. Analysis of Bottlenecks in Experimental Models of Infection. PLoS Pathog. 2015;11:e1004823.Global Burden of Disease Study. 2013 Collaborators. Global, Regional, and National Incidence, Prevalence, and Years Lived with Disability for 301 Acute and Chronic Diseases and Injuries in 188 Countries, 1990-2013: A Systematic Analysis for the Global Burden of Disease Study 2013. Lancet Lond Engl 2015;386:743–800.Article 

Google Scholar 
WHO | Diarrhoeal Disease [Internet]. WHO. [cited 2017 Oct 22]. Available from: http://www.who.int/mediacentre/factsheets/fs330/en/.Crepin VF, Collins JW, Habibzay M, Frankel G. Citrobacter rodentium Mouse Model of Bacterial Infection. Nat Protoc. 2016;11:1851–76.CAS 
Article 

Google Scholar 
Bhinder G, Sham HP, Chan JM, Morampudi V, Jacobson K, Vallance BA. The Citrobacter rodentium Mouse Model: Studying Pathogen and Host Contributions to Infectious Colitis. J Vis Exp. 2013;19:e50222.
Google Scholar 
Barthold SW, Coleman GL, Bhatt PN, Osbaldiston GW, Jonas AM. The Etiology of Transmissible Murine Colonic Hyperplasia. Lab Anim Sci. 1976;26:889–94.CAS 
PubMed 

Google Scholar 
Deng W, Vallance BA, Li Y, Puente JL, Finlay BB. Citrobacter rodentium Translocated Intimin Receptor (Tir) is an Essential Virulence Factor Needed for Actin Condensation, Intestinal Colonization and Colonic Hyperplasia in Mice. Mol Microbiol. 2003;48:95–115.CAS 
Article 

Google Scholar 
Schauer DB, Falkow S. Attaching and Effacing Locus of a Citrobacter freundii Biotype That Causes Transmissible Murine Colonic Hyperplasia. Infect Immun. 1993;61:2486–92.CAS 
Article 

Google Scholar 
Elliott SJ, Yu J, Kaper JB. The Cloned Locus of Enterocyte Effacement From Enterohemorrhagic Escherichia coli O157:H7 is Unable to Confer the Attaching and Effacing Phenotype Upon E. coli K-12. Infect Immun. 1999;67:4260–3.CAS 
Article 

Google Scholar 
Mellies JL, Elliott SJ, Sperandio V, Donnenberg MS, Kaper JB. The Per Regulon of Enteropathogenic Escherichia coli: Identification of a Regulatory Cascade and a Novel Transcriptional Activator, the Locus of Enterocyte Effacement (LEE)-Encoded Regulator (Ler). Mol Microbiol. 1999;33:296–306.CAS 
Article 

Google Scholar 
Jarvis KG, Girón JA, Jerse AE, McDaniel TK, Donnenberg MS, Kaper JB. Enteropathogenic Escherichia coli Contains a Putative Type III Secretion System Necessary for the Export of Proteins Involved in Attaching and Effacing Lesion Formation. Proc Natl Acad Sci. 1995;92:7996–8000.CAS 
Article 

Google Scholar 
Moon HW, Whipp SC, Argenzio RA, Levine MM, Giannella RA. Attaching and Effacing Activities of Rabbit and Human Enteropathogenic Escherichia coli in Pig and Rabbit Intestines. Infect Immun. 1983;41:1340–51.CAS 
Article 

Google Scholar 
Abel S, Abel zur Wiesch P, Chang HH, Davis BM, Lipsitch M, Waldor MK. Sequence Tag–based Analysis of Microbial Population Dynamics. Nat Methods. 2015;12:223–6.CAS 
Article 

Google Scholar 
Cavalli-Sforza LL, Edwards AW. Phylogenetic analysis. Models and Estimation Procedures. Am J Hum Genet. 1967;19:233–57. 3.CAS 
PubMed 
PubMed Central 

Google Scholar 
Zhang T, Abel S, Abel Zur Wiesch P, Sasabe J, Davis BM, Higgins DE, et al. Deciphering the Landscape of Host Barriers to Listeria monocytogenes Infection. Proc Natl Acad Sci . 2017;114:6334–9.CAS 
Article 

Google Scholar 
Hullahalli K, Waldor MK. Pathogen Clonal Expansion Underlies Multiorgan Dissemination and Organ-Specific Outcomes During Murine Systemic Infection. eLife 2021;10:e70910.CAS 
Article 

Google Scholar 
Petty NK, Feltwell T, Pickard D, Clare S, Toribio AL, Fookes M, et al. Citrobacter rodentium is an Unstable Pathogen Showing Evidence of Significant Genomic Flux. PLOS Pathog. 2011;7:e1002018.CAS 
Article 

Google Scholar 
Yasutomi E, Hoshi N, Adachi S, Otsuka T, Kong L, Ku Y, et al. Proton Pump Inhibitors Increase the Susceptibility of Mice to Oral Infection with Enteropathogenic Bacteria. Dig Dis Sci. 2018;63:881–9.CAS 
Article 

Google Scholar 
Callahan BJ, McMurdie PJ, Rosen MJ, Han AW, Johnson AJA, Holmes SP. DADA2: High-Resolution Sample Inference From Illumina Amplicon Data. Nat Methods. 2016;13:581–3.CAS 
Article 

Google Scholar 
Krimbas CB, Tsakas S. The Genetics of Dacus Oleae. v. Changes of Esterase Polymorphism in a Natural Population Following Insecticide Control-Selection or Drift? Evol Int J Org Evol. 1971;25:454–60.Article 

Google Scholar 
Peter BM, Slatkin M. Detecting Range Expansions From Genetic Data. Evol Int J Org Evol. 2013;67:3274–89.Article 

Google Scholar 
Schwarz R, Kaspar A, Seelig J, Künnecke B. Gastrointestinal Transit Times in Mice and Humans Measured with 27Al and 19F Nuclear Magnetic Resonance. Magn Reson Med. 2002;48:255–61.Article 

Google Scholar 
McMurdie PJ, Holmes S. phyloseq: An R Package for Reproducible Interactive Analysis and Graphics of Microbiome Census Data. PloS One. 2013;8:e61217.CAS 
Article 

Google Scholar 
Petty NK, Bulgin R, Crepin VF, Cerdeño-Tárraga AM, Schroeder GN, Quail MA, et al. The Citrobacter rodentium Genome Sequence Reveals Convergent Evolution with Human Pathogenic Escherichia coli. J Bacteriol. 2010;192:525–38.CAS 
Article 

Google Scholar 
Mundy R, Pickard D, Wilson RK, Simmons CP, Dougan G, Frankel G. Identification of a Novel Type IV Pilus Gene Cluster Required for Gastrointestinal Colonization of Citrobacter rodentium. Mol Microbiol. 2003;48:795–809.CAS 
Article 

Google Scholar 
Darwin AJ, Miller VL. Identification of Yersinia enterocolitica Genes Affecting Survival in an Animal Host Using Signature-Tagged Transposon Mutagenesis. Mol Microbiol. 1999;32:51–62.CAS 
Article 

Google Scholar 
Maroncle N, Balestrino D, Rich C, Forestier C. Identification of Klebsiella pneumoniae Genes Involved in Intestinal Colonization and Adhesion Using Signature-Tagged Mutagenesis. Infect Immun. 2002;70:4729–34.CAS 
Article 

Google Scholar 
Wiles S, Dougan G, Frankel G. Emergence of a ‘Hyperinfectious’ Bacterial State After Passage of Citrobacter rodentium Through the Host Gastrointestinal Tract. Cell Microbiol. 2005;7:1163–72.CAS 
Article 

Google Scholar 
Wiles S, Clare S, Harker J, Huett A, Young D, Dougan G, et al. Organ Specificity, Colonization and Clearance Dynamics In Vivo Following Oral Challenges with the Murine Pathogen Citrobacter rodentium. Cell Microbiol. 2004;6:963–72.CAS 
Article 

Google Scholar 
Kitamoto S, Nagao-Kitamoto H, Kuffa P, Kamada N. Regulation of Virulence: The Rise and Fall of Gastrointestinal Pathogens. J Gastroenterol. 2016;51:195–205.Article 

Google Scholar 
Takumi K, de Jonge R, Havelaar A. Modelling Inactivation of Escherichia coli by Low pH: Application to Passage Through the Stomach of Young and Elderly People. J Appl Microbiol. 2000;89:935–43.CAS 
Article 

Google Scholar 
Pienaar JA, Singh A, Barnard TG. Acid-Happy: Survival and Recovery of Enteropathogenic Escherichia coli (EPEC) in Simulated Gastric Fluid. Micro Pathog. 2019;128:396–404.CAS 
Article 

Google Scholar 
McConnell EL, Basit AW, Murdan S. Measurements of Rat and Mouse Gastrointestinal pH, Fluid and Lymphoid Tissue, and Implications for In-Vivo Experiments. J Pharm Pharm. 2008;60:63–70.CAS 
Article 

Google Scholar 
Tan A, Petty NK, Hocking D, Bennett-Wood V, Wakefield M, Praszkier J, et al. Evolutionary Adaptation of an AraC-Like Regulatory Protein in Citrobacter rodentium and Escherichia Species. Infect Immun. 2015;83:1384–95.CAS 
Article 

Google Scholar 
Sanchez KK, Chen GY, Schieber AMP, Redford SE, Shokhirev MN, Leblanc M, et al. Cooperative Metabolic Adaptations in the Host Can Favor Asymptomatic Infection and Select for Attenuated Virulence in an Enteric Pathogen. Cell 2018;175:146–.e15.CAS 
Article 

Google Scholar 
Cunningham R, Jones L, Enki DG, Tischhauser R. Proton Pump Inhibitor Use as a Risk Factor for Enterobacteriaceal Infection: A Case-Control Study. J Hosp Infect. 2018;100:60–4.CAS 
Article 

Google Scholar 
Kelly OB, Dillane C, Patchett SE, Harewood GC, Murray FE. The Inappropriate Prescription of Oral Proton Pump Inhibitors in the Hospital Setting: A Prospective Cross-Sectional Study. Dig Dis Sci. 2015;60:2280–6.CAS 
Article 

Google Scholar 
Federici S, Suez J, Elinav E. Our Microbiome: On the Challenges, Promises, and Hype. Results Probl Cell Differ. 2020;69:539–57.CAS 
Article 

Google Scholar  More

Gardner, T. A., Côté, I. M., Gill, J. A., Grant, A. & Watkinson, A. R. Long-term region-wide declines in Caribbean corals. Science 301, 958–960 (2003).ADS 
PubMed 
Article 

Google Scholar 
Bruno, J. F. & Selig, E. R. Regional decline of coral cover in the Indo-Pacific: Timing, extent, and subregional comparisons. PLoS ONE 2, e711 (2007).ADS 
PubMed 
PubMed Central 
Article 

Google Scholar 
De’Ath, G., Fabricius, K. E., Sweatman, H. & Puotinen, M. The 27-year decline of coral cover on the Great Barrier Reef and its causes. Proc. Natl. Acad. Sci. U. S. A. 109, 17995–17999 (2012).ADS 
PubMed 
PubMed Central 
Article 

Google Scholar 
Hughes, T. P., Graham, N. A. J., Jackson, J. B. C., Mumby, P. J. & Steneck, R. S. Rising to the challenge of sustaining coral reef resilience. Trends Ecol. Evol. 25, 633–642 (2010).PubMed 
Article 

Google Scholar 
Pandolfi, J. M., Connolly, S. R., Marshall, D. J. & Cohen, A. L. Projecting coral reef futures under global warming and ocean acidification. Science 333, 418–422 (2011).ADS 
PubMed 
Article 

Google Scholar 
Hughes, T. P. et al. Global warming and recurrent mass bleaching of corals. Nature 543, 373–377 (2017).ADS 
PubMed 
Article 

Google Scholar 
Bindoff, N. L. et al. Chapter 5: Changing Ocean, Marine Ecosystems, and Dependent Communities. IPCC Special Report on the Ocean and Cryosphere in a Changing Climate (2019).Richmond, R. H. Reproduction and recruitment in corals: Critical links in the persistence of reefs. In Life and Death of Coral Reefs (ed. Birkeland, C. E.) 175–197 (Springer, 1997).Chapter 

Google Scholar 
Trapon, M. L., Pratchett, M. S., Hoey, A. S. & Baird, A. H. Influence of fish grazing and sedimentation on the early post-settlement survival of the tabular coral Acropora cytherea. Coral Reefs 32, 1051–1059 (2013).ADS 
Article 

Google Scholar 
Gallagher, C. & Doropoulos, C. Spatial refugia mediate juvenile coral survival during coral–predator interactions. Coral Reefs 36, 51–61 (2017).ADS 
Article 

Google Scholar 
Vermeij, M. J. A. & Sandin, S. A. Density-dependent settlement and mortality structure the earliest life phases of a coral population. Ecology 89, 1994–2004 (2008).PubMed 
Article 

Google Scholar 
Vermeij, M. J. A., Smith, J. E., Smith, C. M., Vega Thurber, R. & Sandin, S. A. Survival and settlement success of coral planulae: Independent and synergistic effects of macroalgae and microbes. Oecologia 159, 325–336 (2009).ADS 
PubMed 
Article 

Google Scholar 
Ricardo, G. F., Jones, R. J., Nordborg, M. & Negri, A. P. Settlement patterns of the coral Acropora millepora on sediment-laden surfaces. Sci. Total Environ. 609, 277–288 (2017).ADS 
PubMed 
Article 

Google Scholar 
Brunner, C. A., Uthicke, S., Ricardo, G. F., Hoogenboom, M. O. & Negri, A. P. Climate change doubles sedimentation-induced coral recruit mortality. Sci. Total Environ. 768, 143897 (2021).ADS 
PubMed 
Article 

Google Scholar 
Birrell, C. L., McCook, L. J., Willis, B. L. & Diaz-Pulido, G. A. Effects of benthic algae on the replenishment of corals and the implications for the resilience of coral reefs. In Oceanography and Marine Biology: An Annual Review 25–63 (CRC Press, 2008).Chapter 

Google Scholar 
Karcher, D. B. et al. Nitrogen eutrophication particularly promotes turf algae in coral reefs of the central Red Sea. PeerJ 2020, 1–25 (2020).
Google Scholar 
Kirschner, C. M. & Brennan, A. B. Bio-inspired antifouling strategies. Annu. Rev. Mater. Res. 42, 211–229 (2012).ADS 
Article 

Google Scholar 
Webster, N. S. et al. Metamorphosis of a scleractinian coral in response to microbial biofilms. Appl. Environ. Microbiol. 70, 1213–1221 (2004).ADS 
PubMed 
PubMed Central 
Article 

Google Scholar 
Heyward, A. J. & Negri, A. P. Natural inducers for coral larval metamorphosis. Coral Reefs 18, 273–279 (1999).Article 

Google Scholar 
Negri, A. P., Webster, N. S., Hill, R. T. & Heyward, A. J. Metamorphosis of broadcast spawning corals in response to bacteria isolated from crustose algae. Mar. Ecol. Prog. Ser. 223, 121–131 (2001).ADS 
Article 

Google Scholar 
Tebben, J. et al. Induction of larval metamorphosis of the coral Acropora millepora by tetrabromopyrrole isolated from a Pseudoalteromonas bacterium. PLoS ONE 6, 1–8 (2011).Article 

Google Scholar 
Sneed, J. M., Sharp, K. H., Ritchie, K. B. & Paul, V. J. The chemical cue tetrabromopyrrole from a biofilm bacterium induces settlement of multiple Caribbean corals. Proc. R. Soc. B Biol. Sci. 281, 1–9 (2014).
Google Scholar 
Tebben, J. et al. Chemical mediation of coral larval settlement by crustose coralline algae. Sci. Rep. 5, 1–11 (2015).Article 

Google Scholar 
Carpenter, R. C. & Edmunds, P. J. Local and regional scale recovery of Diadema promotes recruitment of scleractinian corals. Ecol. Lett. 9, 268–277 (2006).Article 

Google Scholar 
Box, S. J. & Mumby, P. J. Effect of macroalgal competition on growth and survival of juvenile Caribbean corals. Mar. Ecol. Prog. Ser. 342, 139–149 (2007).ADS 
Article 

Google Scholar 
Linares, C., Cebrian, E. & Coma, R. Effects of turf algae on recruitment and juvenile survival of gorgonian corals. Mar. Ecol. Prog. Ser. 452, 81–88 (2012).ADS 
Article 

Google Scholar 
McCook, L. J., Jompa, J. & Diaz-Pulido, G. Competition between corals and algae on coral reefs: A review of evidence and mechanisms. Coral Reefs 19, 400–417 (2001).ADS 
Article 

Google Scholar 
Nugues, M. M., Smith, G. W., Van Hooidonk, R. J., Seabra, M. I. & Bak, R. P. M. Algal contact as a trigger for coral disease. Ecol. Lett. 7, 919–923 (2004).Article 

Google Scholar 
Fong, J. et al. Allelopathic effects of macroalgae on Pocillopora acuta coral larvae. Mar. Environ. Res. 151, 104745. https://doi.org/10.1016/j.marenvres.2019.06.007 (2019).Article 
PubMed 

Google Scholar 
Hauri, C., Fabricius, K. E., Schaffelke, B. & Humphrey, C. Chemical and physical environmental conditions underneath mat- and canopy-forming macroalgae, and their effects on understorey corals. PLoS ONE 5, 1–9 (2010).Article 

Google Scholar 
Bay, L. K. et al. Reef Restoration and Adaptation Program : Intervention Technical Summary. A report provided to the Australian Government by the Reef Restoration and Adaptation Program. (2019).Anthony, K. R. N. et al. Interventions to help coral reefs under global change—A complex decision challenge. PLoS ONE 15, 1–14 (2020).Article 

Google Scholar 
Vardi, T. et al. Six priorities to advance the science and practice of coral reef restoration worldwide. Restor. Ecol. 29, 1–7 (2021).Article 

Google Scholar 
Heyward, A. J., Rees, M. & Smith, L. D. Coral spawning slicks harnessed for large-scale coral culture. Progr. Abstr. Int. Conf. Sci. Asp. Coral Reef Assess. Monit. Restor. 104, 188–189 (1999).
Google Scholar 
Harrison, P., Villanueva, R. & De la Cruz, D. Coral Reef Restoration using Mass Coral Larval Reseeding (Southern Cross University, 2016).
Google Scholar 
de la Cruz, D. W. & Harrison, P. L. Enhancing coral recruitment through assisted mass settlement of cultured coral larvae. PLoS ONE 15, e0242847. https://doi.org/10.1371/journal.pone.0242847 (2020).Article 

Google Scholar 
Chamberland, V. F., Snowden, S., Marhaver, K. L., Petersen, D. & Vermeij, M. J. A. The reproductive biology and early life ecology of a common Caribbean brain coral, Diploria labyrinthiformis (Scleractinia: Faviinae). Coral Reefs 36, 83–94 (2017).ADS 
Article 

Google Scholar 
Randall, C. J. et al. Sexual production of corals for reef restoration in the Anthropocene. Mar. Ecol. Prog. Ser. 635, 203–232 (2020).ADS 
Article 

Google Scholar 
Miller, M. W. et al. Settlement yields in large-scale in situ culture of Caribbean coral larvae for restoration. Restor. Ecol. https://doi.org/10.1111/rec.13512 (2021).Article 

Google Scholar 
Baria-Rodriguez, M. V., de la Cruz, D. W., Dizon, R. M., Yap, H. T. & Villanueva, R. D. Performance and cost-effectiveness of sexually produced Acropora granulosa juveniles compared with asexually generated coral fragments in restoring degraded reef areas. Aquat. Conserv. Mar. Freshw. Ecosyst. 29, 891–900 (2019).Article 

Google Scholar 
Doropoulos, C., Elzinga, J., ter Hofstede, R., van Koningsveld, M. & Babcock, R. C. Optimizing industrial-scale coral reef restoration: Comparing harvesting wild coral spawn slicks and transplanting gravid adult colonies. Restor. Ecol. 27, 758–767 (2019).Article 

Google Scholar 
Kuffner, I. B., Andersson, A. J., Jokiel, P. L., Rodgers, K. S. & MacKenzie, F. T. Decreased abundance of crustose coralline algae due to ocean acidification. Nat. Geosci. 1, 114–117 (2008).ADS 
Article 

Google Scholar 
Webster, N. S., Uthicke, S., Botté, E. S., Flores, F. & Negri, A. P. Ocean acidification reduces induction of coral settlement by crustose coralline algae. Glob. Change Biol. 19, 303–315 (2013).ADS 
Article 

Google Scholar 
Randall, C. J., Giuliano, C., Heyward, A. J. & Negri, A. P. Enhancing coral survival on deployment devices with microrefugia. Front. Mar. Sci. 8, 662263. https://doi.org/10.3389/fmars.2021.662263 (2021).Article 

Google Scholar 
Kuffner, I. B. et al. Inhibition of coral recruitment by macroalgae and cyanobacteria. Mar. Ecol. Prog. Ser. 323, 107–117 (2006).ADS 
Article 

Google Scholar 
Arnold, S. N., Steneck, R. S. & Mumby, P. J. Running the gauntlet: Inhibitory effects of algal turfs on the processes of coral recruitment. Mar. Ecol. Prog. Ser. 414, 91–105 (2010).ADS 
Article 

Google Scholar 
Speare, K. E., Duran, A., Miller, M. W. & Burkepile, D. E. Sediment associated with algal turfs inhibits the settlement of two endangered coral species. Mar. Pollut. Bull. 144, 189–195 (2019).PubMed 
Article 

Google Scholar 
Tebben, J., Guest, J. R., Sin, T. M., Steinberg, P. D. & Harder, T. Corals like it waxed: Paraffin-based antifouling technology enhances coral spat survival. PLoS ONE 9, 1–8 (2014).Article 

Google Scholar 
Almeida, E., Diamantino, T. C. & de Sousa, O. Marine paints: The particular case of antifouling paints. Prog. Org. Coat. 59, 2–20 (2007).Article 

Google Scholar 
Negri, A. P., Smith, L. D., Webster, N. S. & Heyward, A. J. Understanding ship-grounding impacts on a coral reef: Potential effects of anti-foulant paint contamination on coral recruitment. Mar. Pollut. Bull. 44, 111–117 (2002).PubMed 
Article 

Google Scholar 
Smith, L. D., Negri, A. P., Philipp, E., Webster, N. S. & Heyward, A. J. The effects of antifoulant-paint-contaminated sediments on coral recruits and branchlets. Mar. Biol. 143, 651–657 (2003).Article 

Google Scholar 
Jacobson, A. H. & Willingham, G. L. Sea-nine antifoulant: An environmentally acceptable alternative to organotin antifoulants. Sci. Total Environ. 258, 103–110 (2000).ADS 
PubMed 
Article 

Google Scholar 
Silva, V. et al. Isothiazolinone biocides: Chemistry, biological, and toxicity profiles. Molecules 25, 991. https://doi.org/10.3390/molecules25040991 (2020).Article 
PubMed Central 

Google Scholar 
da Silva, A. R., da Guerreiro, A. S., Martins, S. E. & Sandrini, J. Z. DCOIT unbalances the antioxidant defense system in juvenile and adults of the marine bivalve Amarilladesma mactroides (Mollusca: Bivalvia). Comp. Biochem. Physiol. Part C 250, 109169 (2021).
Google Scholar 
Cima, F. et al. Preliminary evaluation of the toxic effects of the antifouling biocide Sea-Nine 211TM in the soft coral Sarcophyton cf. glaucum (Octocorallia, Alcyonacea) based on PAM fluorometry and biomarkers. Mar. Environ. Res. 83, 16–22 (2013).PubMed 
Article 

Google Scholar 
Wendt, I., Backhaus, T., Blanck, H. & Arrhenius, Å. The toxicity of the three antifouling biocides DCOIT, TPBP and medetomidine to the marine pelagic copepod Acartia tonsa. Ecotoxicology 25, 871–879 (2016).PubMed 
Article 

Google Scholar 
Chen, L. et al. Identification of molecular targets for 4,5-dichloro-2-n-octyl-4-isothiazolin-3-one (DCOIT) in teleosts: New insight into mechanism of toxicity. Environ. Sci. Technol. 51, 1840–1847 (2017).ADS 
PubMed 
Article 

Google Scholar 
Martins, S. E., Fillmann, G., Lillicrap, A. & Thomas, K. V. Review: Ecotoxicity of organic and organo-metallic antifouling co-biocides and implications for environmental hazard and risk assessments in aquatic ecosystems. Biofouling 34, 34–52 (2018).PubMed 
Article 

Google Scholar 
Moon, Y. S., Kim, M., Hong, C. P., Kang, J. H. & Jung, J. H. Overlapping and unique toxic effects of three alternative antifouling biocides (Diuron, Irgarol 1051 ®, Sea-Nine 211 ® ) on non-target marine fish. Ecotoxicol. Environ. Saf. 180, 23–32 (2019).PubMed 
Article 

Google Scholar 
Su, Y. et al. Toxicity of 4,5-dichloro-2-n-octyl-4-isothiazolin-3-one (DCOIT) in the marine decapod Litopenaeus vannamei. Environ. Pollut. 251, 708–716 (2019).PubMed 
Article 

Google Scholar 
Fonseca, V. B., da Guerreiro, A. S., Vargas, M. A. & Sandrini, J. Z. Effects of DCOIT (4,5-dichloro-2-octyl-4-isothiazolin-3-one) to the haemocytes of mussels Perna perna. Comp. Biochem. Physiol Part C 232, 108737. https://doi.org/10.1016/j.cbpc.2020.108737 (2020).Article 

Google Scholar 
Ferreira, V. et al. Effects of nanostructure antifouling biocides towards a coral species in the context of global changes. Sci. Total Environ. 799, 149324 (2021).ADS 
PubMed 
Article 

Google Scholar 
de Campos, B. G. et al. A preliminary study on multi-level biomarkers response of the tropical oyster Crassostrea brasiliana to exposure to the antifouling biocide DCOIT. Mar. Pollut. Bull. 174, 112141 (2022).Article 

Google Scholar 
Maia, F. et al. Incorporation of biocides in nanocapsules for protective coatings used in maritime applications. Chem. Eng. J. 270, 150–157 (2015).Article 

Google Scholar 
Santos, J. V. N. et al. Can encapsulation of the biocide DCOIT affect the anti-fouling efficacy and toxicity on tropical bivalves?. Appl. Sci. 10, 1–12 (2020).Article 

Google Scholar 
Detty, M. R., Ciriminna, R., Bright, F. V. & Pagliaro, M. Environmentally benign sol-gel antifouling and foul-releasing coatings. Acc. Chem. Res. 47, 678–687 (2014).PubMed 
Article 

Google Scholar 
Korschelt, K., Tahir, M. N. & Tremel, W. A Step into the future: Applications of nanoparticle enzyme mimics. Chemistry 24, 9703–9713 (2018).PubMed 
Article 

Google Scholar 
Herget, K. et al. Haloperoxidase mimicry by CeO2-x nanorods combats biofouling. Adv. Mater. 29, 1–8 (2017).Article 

Google Scholar 
Korschelt, K. et al. CeO2-: X nanorods with intrinsic urease-like activity. Nanoscale 10, 13074–13082 (2018).PubMed 
Article 

Google Scholar 
Herget, K., Frerichs, H., Pfitzner, F., Tahir, M. N. & Tremel, W. Functional enzyme mimics for oxidative halogenation reactions that combat biofilm formation. Adv. Mater. 30, 1–28 (2018).Article 

Google Scholar 
Doropoulos, C., Ward, S., Marshell, A., Diaz-Pulido, G. & Mumby, P. J. Interactions among chronic and acute impacts on coral recruits: The importance of size-escape thresholds. Ecology 93, 2131–2138 (2012).PubMed 
Article 

Google Scholar 
Ji, Z. et al. Designed synthesis of CeO2 nanorods and nanowires for studying toxicological effects of high aspect ratio nanomaterials. ACS Nano 6, 5366–5380 (2012).PubMed 
PubMed Central 
Article 

Google Scholar 
Herget, K. et al. Supporting Information: Haloperoxidase mimicry by CeO2-x nanorods combats biofouling. Adv. Mater. 29, 1603823 (2017).Article 

Google Scholar 
Sokolova, A. et al. Spontaneous multiscale phase separation within fluorinated xerogel coatings for fouling-release surfaces. Biofouling 28, 143–157 (2012).PubMed 
Article 

Google Scholar 
Schneider, C. A., Rasband, W. S. & Eliceiri, K. W. NIH Image to ImageJ: 25 years of image analysis. Nat. Methods 9, 671–675 (2012).PubMed 
PubMed Central 
Article 

Google Scholar 
ImageJ Release Notes. https://imagej.nih.gov/ij/notes.html.Arganda-Carreras, I. et al. Trainable Weka Segmentation: A machine learning tool for microscopy pixel classification. Bioinformatics 33, 2424–2426 (2017).PubMed 
Article 

Google Scholar 
Arganda-Carreras, I. et al. Supplementary Data: Trainable Weka Segmentation: A Machine Learning Tool for Microscopy Pixel Classification: Trainable Weka Segmentation User Manualhttps://doi.org/10.1093/bioinformatics/btx180 (2017).Vyas, N., Sammons, R. L., Addison, O., Dehghani, H. & Walmsley, A. D. A quantitative method to measure biofilm removal efficiency from complex biomaterial surfaces using SEM and image analysis. Sci. Rep. 6, 2–11 (2016).Article 

Google Scholar 
Carbone, D. A., Gargano, I., Pinto, G., De Natale, A. & Pollio, A. Evaluating microalgae attachment to surfaces: A first approach towards a laboratory integrated assessment. Chem. Eng. Trans. 57, 73–78 (2017).
Google Scholar 
Moreno Osorio, J. H. et al. Early colonization stages of fabric carriers by two Chlorella strains. J. Appl. Phycol. 32, 3631–3644 (2020).Article 

Google Scholar 
Ricardo, G. F. et al. Impacts of water quality on Acropora coral settlement: The relative importance of substrate quality and light. Sci. Total Environ. 777, 146079. https://doi.org/10.1016/j.scitotenv.2021.146079 (2021).ADS 
Article 
PubMed 

Google Scholar 
Macadam, A., Nowell, C. J. & Quigley, K. Machine learning for the fast and accurate assessment of fitness in coral early life history. Remote Sens. 13, 1–17 (2021).Article 

Google Scholar 
Negri, A. P. & Heyward, A. J. Inhibition of Fertilization and Larval Metamorphosis of the Coral Acropora millepora (Ehrenberg, 1834) by Petroleum Products. Mar. Pollut. Bull. 41, 420–427 (2000).Article 

Google Scholar 
Nordborg, F. M., Flores, F., Brinkman, D. L., Agustí, S. & Negri, A. P. Phototoxic effects of two common marine fuels on the settlement success of the coral Acropora tenuis. Sci. Rep. 8, 1–12 (2018).Article 

Google Scholar 
R Core Team. R: A Language and Environment for Statistical Computing. (2021).Wickham, H. et al. Welcome to the Tidyverse. J. Open Source Softw. 4, 1686. https://doi.org/10.21105/joss.01686 (2019).ADS 
Article 

Google Scholar 
Pinheiro, J., Bates, D., Debroy, S., Sarkar, D. & R Core Team. Linear and nonlinear mixed effects models contact. Linear nonlinear Mix. Eff. Model. 3, 103–135 (2021).
Google Scholar 
Fox, J. & Weisberg, S. An R Companion to Applied Regression (Sage Publications, 2019).
Google Scholar 
Lenth, R. V. Emmeans: Estimated Marginal Means. https://cran.r-project.org/package=emmeans (2021).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 
Dafforn, K. A., Lewis, J. A. & Johnston, E. L. Antifouling strategies: History and regulation, ecological impacts and mitigation. Mar. Pollut. Bull. 62, 453–465 (2011).PubMed 
Article 

Google Scholar 
Wu, R. et al. Room temperature synthesis of defective cerium oxide for efficient marine anti-biofouling. Adv. Compos. Hybrid Mater. https://doi.org/10.1007/s42114-021-00256-7 (2021).Article 

Google Scholar 
Hu, M. et al. Nanozymes in nanofibrous mats with haloperoxidase-like activity to combat biofouling. ACS Appl. Mater. Interfaces 10, 44722–44730 (2018).PubMed 
Article 

Google Scholar 
He, X. et al. Haloperoxidase mimicry by CeO2-x nanorods of different aspect ratios for antibacterial performance. ACS Sustain. Chem. Eng. 8, 6744–6752 (2020).Article 

Google Scholar 
Saxena, P. & Harish,. Nanoecotoxicological reports of engineered metal oxide nanoparticles on algae. Curr. Pollut. Rep. 4, 128–142 (2018).Article 

Google Scholar 
Xu, Y. et al. Effects of cerium oxide nanoparticles on bacterial growth and behaviors: Induction of biofilm formation and stress response. Environ. Sci. Pollut. Res. 26, 9293–9304 (2019).Article 

Google Scholar 
Xu, Y. et al. Mechanistic understanding of cerium oxide nanoparticle-mediated biofilm formation in Pseudomonas aeruginosa. Environ. Sci. Pollut. Res. 25, 34765–34776 (2018).Article 

Google Scholar 
Tang, Y. et al. Hybrid xerogel films as novel coatings for antifouling and fouling release. Biofouling 21, 59–71 (2005).PubMed 
Article 

Google Scholar 
Gunari, N. et al. The control of marine biofouling on xerogel surfaces with nanometer-scale topography. Biofouling 27, 137–149 (2011).PubMed 
Article 

Google Scholar 
Maia, F. et al. Silica nanocontainers for active corrosion protection. Nanoscale 4, 1287–1298 (2012).ADS 
PubMed 
Article 

Google Scholar 
Martins, R. et al. Effects of a novel anticorrosion engineered nanomaterial on the bivalve: Ruditapes philippinarum. Environ. Sci. Nano 4, 1064–1076 (2017).Article 

Google Scholar 
Gutner-Hoch, E. et al. Antimacrofouling efficacy of innovative inorganic nanomaterials loaded with booster biocides. J. Mar. Sci. Eng. 6, 15. https://doi.org/10.3390/jmse6010006 (2018).Article 

Google Scholar 
Negri, A. P. & Heyward, A. J. Inhibition of coral fertilisation and larval metamorphosis by tributyltin and copper. Mar. Environ. Res. 51, 17–27 (2001).PubMed 
Article 

Google Scholar 
Morse, D. E., Hooker, N., Morse, A. N. C. & Jensen, R. A. Control of larval metamorphosis and recruitment in sympatric agariciid corals. J. Exp. Mar. Biol. Ecol. 116, 193–217 (1988).Article 

Google Scholar 
Harrington, L., Fabricius, K., De’ath, G. & Negri, A. Recognition and selection of settlement substrata determine post-settlement survival in corals. Ecology 85, 3428–3437 (2004).Article 

Google Scholar 
Jorissen, H., Baumgartner, C., Steneck, R. S. & Nugues, M. M. Contrasting effects of crustose coralline algae from exposed and subcryptic habitats on coral recruits. Coral Reefs 39, 1767–1778 (2020).Article 

Google Scholar 
Figueiredo, J. et al. Toxicity of innovative anti-fouling nano-based solutions to marine species. Environ. Sci. Nano 6, 1418–1429 (2019).Article 

Google Scholar 
Shafir, S., Abady, S. & Rinkevich, B. Improved sustainable maintenance for mid-water coral nursery by the application of an anti-fouling agent. J. Exp. Mar. Biol. Ecol. 368, 124–128 (2009).Article 

Google Scholar  More

Study site and animalsWe studied n = 64 flocks of Japanese tits in mixed deciduous-coniferous forests in Nagano and Gumma (36°17-31’N, 138°26-39’E), Japan. Although most of the birds had not been individually colour-ringed, all the experimental trials were conducted at least 400 m apart; previous observations on colour-ringed individuals showed that this distance was enough to ensure the collection of data from different individuals30. In this site, one of the major predators of small birds is the bull-headed shrike, which is often mobbed by small birds including Japanese tits.Playback stimulusTo test whether Japanese tits recognize an alert-recruitment call sequence as a single unit, we prepared four treatments: (i) one-speaker playback of alert-recruitment call sequences, (ii) two-speaker playback of alert-recruitment call sequences with alert and recruitment calls played from different speakers, (iii) one-speaker playback of recruitment-alert call sequences, (iv) two-speaker playback of recruitment-alert call sequences with recruitment and alert calls played from different speakers (Fig. 3). We created sound files for these treatments using the software program Audacity 2.1.3 (http://www.audacityteam.org). For one-speaker treatments, we composed mono sound files where call sequences were repeated onto a single channel, whereas for two-speaker treatments, we composed stereo sound files where either alert or recruitment calls were repeated onto the right or left channels, respectively. All the files contained an equal number of alert calls (30 calls) and recruitment calls (30 calls) at the same rate (one call every 3 s), resulting in 90-s of stimuli (Fig. 3), which corresponds to the range of the natural calling rate of alert-recruitment sequences during mobbing by Japanese tits10. For all stimuli, within-call-sequence intervals between alert and recruitment calls were constant (0.1 s), which is within the range of intervals of these calls in natural call sequences11,17. In contrast, between-call-sequence intervals varied from 1.50 to 1.81 (median = 1.68) due to the difference in call length, but were constant across playback stimuli within the same “block” where the four treatments were created using the same call exemplars (see below). While alert calls are composed of three distinct note types, recruitment calls are strings of the same note type that vary in repetition number. Since the repetition number can vary depending on predator type10, we conducted predator exposure experiments to Japanese tit flocks (n = 12) and recorded call sequences towards a bull-headed shrike life-like specimen. In response to a shrike specimen, tits produced alert-recruitment call sequences with a recruitment note repetition number ranging from 5 to 15. Since the interquartile range of repetition number was 6.75 to 10, we used recruitment calls with 7–10 notes as playback stimuli in this study. In consideration for the possible influence of sound editing procedure, we created all the stimuli in the same manner; we copied alert and recruitment call parts separately from recording files, and pasted them onto background noise files to produce all four types of stimuli. Playback amplitudes were constant across treatments, 70 dB at 1.0 m measured using a sound level meter (SM-325, AS ONE Corporation). Therefore, the differences between treatments only depend on whether these calls are produced as sequences from the same source and how the calls are ordered.We carefully designed experiments to control for the possibility that individual-based acoustic features in alert and recruitment calls might influence tits’ responses. First, we prepared 16 unique sets of alert and recruitment calls using either calls from the same bird (n = 8 source individuals, n = 8 unique call sets) or from two different birds (n = 16 source individuals, n = 8 unique call sets). Then, we created the four types of treatments (i.e., alert-recruitment call sequences from the same speaker, from different speakers, and in reversed order from the same speaker and from different speakers) from each of the alert-recruitment call sets, resulting in 16 blocks of playback stimuli (Supplementary Table 3). This allows us to test the possible influence of individual-based acoustic variation on receivers’ responses.We were also careful to avoid the possible influence of population-level signatures of acoustic features: we only used Japanese tits’ call sequences that had been previously recorded from the same study population. We saved the sound files in .wav format (16-bit accuracy, 48-kHz sampling rate) onto a playback device (iPhone 8, Apple Inc.). We used the default Music app (Apple Inc.) to playback the sound files.ExperimentWe (TNS and YKM) conducted experimental trials from 26 October to 4 December 2020 and during the period of 0800 and 1600 h (Japan Standard Time). We did not conduct trials under wet and windy weather conditions, since these may influence behavioural patterns of forest birds31. First, we searched for and located a flock of Japanese tits. Upon finding a flock, we fixed a taxidermic specimen of bull-headed shrike in a perching posture on the branch at 1.8 ± 0.2 m (mean ± s.d., n = 64) above the ground. Then, we placed either one or two Bluetooth speakers (SoundLink Micro, BOSE) on tree branches at 1.6 ± 0.2 m (mean ± s.d., n = 96) above the ground, and oriented them upwards to control for the possible influence of directionality. We set the distance between the shrike specimen and the speaker(s) at 5 m. For trials with two speakers, we set the distance between speakers at 10 m, mimicking the situation in which two birds are calling (Fig. 3). The shrike specimen was first covered with a black cloth and was exposed by removing the cloth just before each trial.We began playbacks when at least two Japanese tits were present within 15 m from the shrike specimen. During 90-s of playbacks, we recorded (i) whether birds approached within 2-m of the shrike specimen during the playback and (ii) whether birds exhibited wing flicking displays12,13. We counted the number of individuals within 15 m from the shrike during 90-s of playbacks and considered it as flock size. During trials, we sat on the ground at ca. 10 m from the shrike specimen to decrease the influence of the observers’ presence on bird behaviour. To account for the inter-observer reliability32, we calculated intra-class correlation coefficient (ICC; icc function in the R package irr) between us. The lowest ICC was 0.998, indicating high degree of inter-observer reliability for the two behavioural measurements. We also video-recorded the responses of tits using a digital video camera (FDR-AX60, SONY). After completion of each trial, we checked the video recording and made an on-the-spot confirmation of the exact location at which each bird made the closest approach to the shrike specimen during the 90-s of playbacks. Then, using a tape measure, we recorded the minimum approach distance of birds to the shrike specimen. Thus, our final data set consisted of the most reliable observations confirmed by two experimenters and video evidence.The order of trials was randomized within each block (n = 16 blocks), each of which is composed of a unique alert-recruitment call set but includes four treatments differing in the number of speakers and call order. Therefore, responses to all four treatments were observed under largely similar conditions. In a few trials, the first bird to approach the shrike specimen was from a heterospecific species, such as a varied tit (n = 1) or a long-tailed tit (n = 1). To account for the possibility that these birds evoke mobbing behaviour in Japanese tits, we only used the data from instances where the first individual to approach the shrike was a Japanese tit. Otherwise, we repeated the same treatment at a different site.We used 64 unique playbacks created from 16 unique sets of alert-recruitment calls for 64 trials in order to avoid pseudoreplication33. We prepared two specimens of male bull-headed shrikes and used each of them for the equal number of trials. We did not use specimens of female shrikes since females migrate from the study site in late summer and only males were observed during the study period.Statistical analysisWe analyzed the effect of playback treatments on the mobbing behaviours of Japanese tits using generalized linear mixed models in R34,35. We used the proportions of Japanese tits in flocks that (i) approached within 2-m of the shrike specimen and (ii) exhibited wing flicking displays. For the analysis of predator approach, we prepared two vectors (i.e., the number of Japanese tits that approached the shrike specimen and the number of Japanese tits that did not approach the shrike specimen). Then, we created a single response variable by binding together these two vectors using cbind function. Similarly, for the analysis of wing flicking displays, we created a single response variable by binding two vectors (i.e., the number of tits that exhibit wing flicking and the number of tits that did not exhibit wing flicking). We fitted playback treatments as a fixed term, and flock size (maximum number of Japanese tits observed during 90-s of playback) and the way of creating playback stimuli (whether the two call types were recorded from a single individual or two individuals) as covariates. We also included identity of alert-recruitment call sets that were used for creating playback stimuli (i.e., call sets from either one or two source individuals) and identity of shrike specimens as random terms. We used a binomial error distribution and logit-link function (glmer in the R package lme4) for these models. Statistical significance was calculated by log-likelihood ratio tests using anova in the R package stats. We further conducted post-hoc pairwise comparisons between treatments by using estimated marginal means (emmeans in the R package emmeans). When making pairwise comparisons, we adjusted p-values by applying a false discovery rate control for multiple testing36. All tests were two-sided and the significance level was set at α = 0.05. Exact p-values are reported when p ≥ 0.0001.Ethics statementAll protocols were approved by the ethics committee of Kyoto University, the Ministry of the Environment, and the Forestry Agency of Japan, and adhered to Guidelines for the Use of Animals of the Association for the Study of Animal Behaviour/Animal Behavior Society37.Reporting summaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article. More

Ethical statementThis research project was ethically reviewed and approved by the Ethics Committee of the Instituto Gulbenkian de Ciência (license reference: A009.2018), and by the Portuguese National Entity that regulates the use of laboratory animals (DGAV – Direção Geral de Alimentação e Veterinária (license reference: 008958). All experiments conducted on animals followed the Portuguese (Decreto-Lei n° 113/2013) and European (Directive 2010/63/EU) legislations, concerning housing, husbandry and animal welfare.Escherichia coli clonesThe ancestral invader E. coli strain expresses a Yellow Fluorescent Protein (YFP), and carries streptomycin and ampicillin resistance markers for easiness of isolation from the mouse feces [galK::amp (pZ12)::PLlacO−1-YFP, strR (rpsl150), ΔlacIZYA::scar]. An E. coli strain used for the in vivo competition experiments is isogenic to the ancestral invader but expresses a Cyan Fluorescent Protein (CFP) and carries streptomycin and chloramphenicol resistance markers [galK::chlor (pZ12)::PLlacO−1-CFP, strR (rpsl150), ΔlacIZYA::scar]. The resident E. coli lineage was isolated from the feces along time using McConkey + 0.4% lactose medium, as previously described9. All the resident clones sampled from each mouse belong to E.coli phylogenetic group B9. The invader E. coli strains (YFP and CFP) derive from the K-12 MG1655 strain (DM08) and exhibit a gat negative phenotype, gatZ::IS112. The resident E. coli clone used for the competition experiments in the mouse gut expresses a mCherry fluorescent protein and a chloramphenicol resistance marker, allowing to distinguish the invader and resident strains in the mice feces.E. coli clones were grown at 37 °C under aeration in liquid media Luria broth (LB) from SIGMA — or McConkey and LB agar plates. Media were supplemented with antibiotics streptomycin (100 µg/mL), ampicillin (100 µg/mL) or chloramphenicol (30 µg/mL) when specified.Serial plating of 1X PBS dilutions of feces in LB agar plates supplemented with the appropriate antibiotics were incubated overnight and YFP, CFP or mCherry-labeled bacterial numbers were assessed by counting the fluorescent colonies using a fluorescent stereoscope (SteREO Lumar, Carl Zeiss). The detection limit for bacterial plating was ~300 CFU/g of feces9.In vivo evolution and competition experimentsAll mice (Mus musculus) used in this study were supplied by the Rodent Facility at Instituto Gulbenkian de Ciência (IGC) and were given ad libitum access to food (Rat and Mouse No.3 Breeding (Special Diets Services) and water. Mice were kept at 20-24 °C and 40-60% humidity with a 12-h light-dark cycle. For the in vivo evolution experiment we used the gut colonization model previously established9. Briefly, mice drank water with streptomycin (5 g/L) only for 24 h before a 4 h starvation period of food and water. The animals were then inoculated by gavage with 100 µL of an E. coli bacterial suspension of ~108 colony-forming units (CFUs). Mice A2, B2, D2, E2, G2, H2 and I2 were successfully colonized with the invader E. coli, while mice C2 and F2 failed to be colonized. Six- to eight-week-old C57BL/6 J non-littermate female mice were kept in individually ventilated cages under specified pathogen-free (SPF) barrier conditions at the IGC animal facility. Fecal pellets were collected during more than one year ( >400 days) and stored in 15% glycerol at −80 °C for later analysis. In the competition experiments between the invader ancestral E. coli and evolved populations, we colonized the mice using a 1:1 ratio of each genotype, with bacterial loads being assessed and frozen on a daily basis after gavage.In vivo competition experiments in which the two modes of selection (directional and diversifying) were acting for a longer time period were performed using evolved invader E. coli populations colonizing mice D2, B2 and A2, H2. Here we used both male (n = 8) and female (n = 8) C57BL/6 J mice aged six- to eight-week-old treated with streptomycin during 3 days before gavage. E. coli populations evolving for short time periods do not allow for strong conclusions on which mode of selection is taking place. Evolved invader populations such as I2 or G2 were therefore not used for in vivo fitness assays. To assess the impact of the mouse resident E. coli in the competitive fitness of dgoR we performed one-to-one competitions between the invader ancestral and dgoR KO clones. We first homogenized the mice microbiotas by co-housing the animals during seven days. The animals (n = 6, female C57BL/6 J mice aged six- to eight-week-old) were then maintained under co-housing and given streptomycin-supplemented (5 g/L) water during seven days to break colonization resistance and eradicate their resident E. coli. At this point, the co-housed mice were removed from the antibiotic-supplemented water for two days. The following day, one group of mice was gavaged with an mCherry-expressing resident E. coli (n = 3 mice) while the other group (n = 3) was not, with all animals being individually caged from this point on and receiving normal water without antibiotic. The day after gavage, all mice were colonized with a mix (1:1) of the invader ancestral and the dgoR KO clones, and the bacterial loads were assessed and frozen on a daily basis.Microbiota analysisFecal DNA was extracted with a QIAamp DNA Stool MiniKit (Qiagen), according to the manufacturer’s instructions and with an additional step of mechanical disruption32. 16 S rRNA gene amplification and sequencing was carried out at the Gene Expression Unit from Instituto Gulbenkian de Ciência, following the service protocol. For each sample, the V4 region of the 16 S rRNA gene was amplified in triplicate, using the primer pair F515/R806, under the following PCR cycling conditions: 94 °C for 3 min, 35 cycles of 94 °C for 60 s, 50 °C for 60 s, and 72 °C for 105 s, with an extension step of 72 °C for 10 min. Samples were then pair-end sequenced on an Illumina MiSeq Benchtop Sequencer, following Illumina recommendations. Sampling for microbiota analysis was performed until the microbiota composition stabilized (~1 year after the antibiotic perturbation).QIIME2 version 2017.1133 was used to analyze the 16 S rRNA sequences by following the authors’ online tutorials (https://docs.qiime2.org/2017.11/tutorials/). Briefly, the demultiplexed sequences were filtered using the “denoise-single” command of DADA2 version 1.1434, and forward and reverse sequences were trimmed in the position in which the 25th percentile’s quality score got below 20. Diversity analysis was performed following the QIIME2 tutorial35. Beta diversity distances were calculated through Unweighted Unifrac36. For taxonomic analysis, OTU were picked by assigning operational taxonomic units at 97% similarity against the Greengenes database version 13 (Greengenes 13_8 99% OTUs (250 bp, V4 region 515 F/806 R))37.Whole-genome sequencing and analysis pipelineDNA was extracted38 from E. coli populations (mixture of  > 1000 clones) or a single clone growing in LB plates supplemented with antibiotic to avoid contamination. DNA concentration and purity were quantified using Qubit and NanoDrop, respectively. The DNA library construction and sequencing were carried out by the IGC genomics facility using the Illumina Miseq platform. Processing of raw reads and variants analysis was based on the previous work39. Briefly, sequencing adapters were removed using fastp version 0.20.040 and raw reads were trimmed bidirectionally by 4 bp window sizes across which an average base quality of 20 was required to be retained. Further retention of reads required a minimum length of 100 bps per read containing at least 50% base pairs with phred scores at or above 20. BBsplit (part of BBMap version 38.9)41 was used to remove likely contaminating reads as explained previously39. Separate reference genomes were used for the alignment of invader (K-12 (substrain MG1655; Accession Number: NC_000913.2)) and resident (Accession Number: SAMN15163749) E. coli genomes. Alignments were performed via three alignment approaches: BWA-sampe version 0.7.1742, MOSAIK version 2.743, and Breseq version 0.35.144,45. Final average alignment depths for invader and resident populations across time points equalled 302 (median = 236) and 253 (median = 235), respectively. While Breseq provides variant analysis in addition to alignment, other variant calling approaches were used to identify putative variation in the sequenced genomes, and to verify data from Breseq. A naïve pipeline39 using the mpileup utility within SAMtools version 1.946 and a custom script written in python was employed. Only reads with a minimum mapping quality of 20 were considered for analysis, and variant calling was limited to bases with call qualities of at least 30. At these positions, a minimum of 5 quality reads had to support a putative variant on both strands (with strand bias, pos. strand / neg. strand, above 0.2 or below 5) for further consideration. Finally, mutations were retained if detected in more than one of the alignment approaches, and if they reached a minimum frequency of 5% at a minimum of one time point sampled. Further simple and complex small variants were considered from freebayes version 0.9.2147 with similar thresholds, while insertion sequence movements and other mobile element activity was inferred via is mapper version 248 and panISa version 0.1.649, as well as Breseq, as previously described39. All putative variants were verified manually in IGV version 2.750,51. Raw sequencing reads were deposited in the sequence read archive under bioproject PRJNA666769. Population dynamics of lineage-specific dynamics and the resulting Muller plots were inferred manually and are meant strictly as a means of presenting the data. In order to generate these plots, mutations were sorted by frequency (descending for each time point at which the population was sampled). The largest frequency mutations were considered major lineages within which minor frequency mutations occurred. Assuming that a mutation, which arises subsequent to a preexisting mutation (an already differentiated lineage) cannot exceed the frequency of that preexisting mutation at any point, and will fluctuate in frequency with the preexisting one, we assigned mutations to the lineages within each population. While this resolved the majority of high frequency and medium frequency mutations, low-frequency mutations within the Muller plots cannot be placed with high confidence, and are only included for completeness.Prophage induction rateTo calculate the maximum prophage induction rate we grew E. coli lysogenic clones, starting with the same initial OD600 values: ~0.1 (Bioscreen C system, Oy Growth Curves Ab Ltd), with agitation at 37 °C in LB medium in the presence or absence of mitomycin C along time (5 µg/mL)9. The OD600 values were normalized by dividing the ones in the presence of mitomycin C by the ones in the absence of mitomycin C (sampling interval: 30 min). The LN of this ratios along time originates a lysis curve, where the maximum slope corresponds to the maximal prophage induction rate for each clone analyzed. We tested evolved clones from mouse A2, H2 and G2 against the ancestral clone which only carries the Nef and the KingRac prophages. We also tested clones of the resident strain that had evolved in the presence of the invader for more than 400 days (these clones were sampled from mouse A2).
E. coli growth rate, growth curves, cell aggregation, biofilm and motility capacityTo calculate the maximum bacterial growth rate, we grew E. coli lysogenic clones, starting with the same initial OD600 values: ~0.1 (Bioscreen C system, Oy Growth Curves Ab Ltd), with agitation at 37 °C in LB medium along time using reading intervals of 30 min. The LN of the OD600 values along time originates a growth curve, where the maximum slope corresponds to the maximum bacterial growth rate for each clone analyzed.To test for metabolic differences of the psuK/fruA mutation, growth curves of evolved lysogenic E. coli clones, bearing the Nef and KingRac prophages, with or without the psuK/fruA mutation were performed with the same initial OD600 value (~0.03) for each clone. The clones were grown in glucose (0.4%) minimal medium (MM9-SIGMA) with or without pseudouridine (80 μM) and absorbance values were obtained using the Bioscreen C apparatus during 12 h.Frozen stocks of E. coli clones were used to seed tubes with 5 mL of liquid LB. These were incubated overnight at 37 °C under static conditions to assess the formation of cell flocks/clumps, observable to the naked eye, in order to evaluate the formation of cell aggregates. Biofilm was tested according a previously published protocol52 and to evaluate the motility capacity we adapted the protocol from Croze and colleagues53. Briefly, overnight E. coli clonal cultures grown with agitation at 37 °C in 5 mL LB medium supplemented with streptomycin (100 ug/mL) were adjusted to the same absorbance and a 3uL volume was dropped on top of soft agar (0.25%). Plates were incubated at 37 °C and photos were taken at day 1, 2 and 5 post-inoculation to assess swarming motility phenotype.Number of E. coli generations during mouse gut colonizationTo estimate the number of generations of E. coli in the mouse gut, we used a previously described protocol to measure the fluorescent intensity of a probe specific to E. coli 23 S rRNA (as a measure of ribosomal content) that correlates with the growth rate of the bacterial cells54. We measured the number of generations of the ancestral E. coli clone while colonizing the gut of 2 mice, treated during 24 h with streptomycin (5 g/L) before gavage, during 25 days.Plasmid DNA extraction and PCR detection of ~69Kb (repA) and ~109Kb (repB) plasmidsPlasmid DNA was extracted from overnight cultures using a Plasmid Mini Kit (Qiagen), according to the manufacturer’s guidelines. Specific primers for the amplification of repA and repB genes, were used to determine the frequency of the 68935 bp (~69 Kb) and 108557 bp (~109 Kb) plasmids, respectively, in the invader E. coli population.The primers used for repA gene were:repA-Forward: 5’-CAGTCCCCTAAAGAATCGCCCC-3’ and repA-Reverse: 5’-TGACCAGGAGCGGCACAATCGC-3’.For repB the primer sequences were:repB-Forward: 5’-GTGGATAAGTCGTCCGGTGAGC-3’ and repB-Reverse: 5’-GTTCAAACAGGCGGGGATCGGC3’.PCR amplification of plasmid-specific genes was performed in 12 isolated random clones from mouse A2 at days 104 and 493. PCR reactions were performed in a total volume of 25 μL, containing 1 μL of plasmid DNA, 1X Taq polymerase buffer, 200 μM dNTPs, 0.2 μM of each primer and 1.25 U Taq polymerase. PCR reaction conditions: 95 °C for 3 min, followed by 35 cycles of 95 °C for 30 s, 65 °C for 30 s and 72 °C for 30 s, finalizing with 5 min at 72 °C. DNA was visualized on a 2% agarose gel stained with GelRed and run at 160 V for 60 min.Construction of the dgoR KO mutantP1 transduction was used to construct a ΔdgoR mutant (dgoR KO). This KO strain was created by replacing the wild-type dgoR in the invader ancestral YFP-expressing genetic background by the respective knock-out from the KEIO collection, strain JW562755, in which the dgoR sequence is replaced by a kanamycin resistance cassette. The presence of the cassette was confirmed by PCR using primers dgoK-F: GCGATGTAGCGAGCTGTC, and yidX-R: GGGAATAAACCGGCAGCC. PCR reactions were performed in a total volume of 25 μL, containing 1 μL of DNA, 1X Taq polymerase buffer, 200 μM dNTPs, 0.2 μM of each primer and 1.25 U Taq polymerase. PCR reaction conditions: 95 °C for 3 min, followed by 35 cycles of 95 °C for 30 s, 65 °C for 30 s and 72 °C for 30 s, finalizing with 5 min at 72 °C. DNA was visualized in a 2% agarose gel stained with GelRed and run at 160 V for 60 min.RNA extraction, DNAse treatment, RT-PCR and qPCRThe Qiagen RNeasy Mini Kit was used for RNA extraction. RNA concentration and quality were evaluated in the Nanodrop 2000 and by gel-electrophoresis. DNase treatment was performed with the RQ1 DNase (Promega) by adding 0.5 μl of DNase to 1 μg of RNA and 1 μl buffer in a final volume of 15 ul, followed by incubation 30 min at 37 °C. Afterwards, 1 ul of stop solution was added and incubation for 15 min at 65 °C was performed to inactivate the DNase. As a control for complete DNA digest a PCR was performed on the reactions including positive controls. Reverse transcription was performed with M-MLV RT[-H] (Promega) by mixing 1 μg of RNA with 0.5 μl random primers (Promega) and nuclease free water to a volume of 15 μl, incubation at 70 °C for 5 min and a quick cool down on ice. Afterwards the reverse transcription was accomplished by adding 5 μl of RT buffer, 0.5 μl RT enzyme and 2 μl dNTP mix, followed by incubation for 10 min at 25 °C, 50 min at 50 °C and 10 min at 70 °C. The resulting cDNA was diluted 100-fold in nuclease free water before changes in gene expression were detected using the The QuantStudio 7Flex (Applied Biosystems) with iTaq Universal SYBR Green Supermix (BioRad) and the following cycling protocol: Hold stage: 2 min at 50 °C, 10 min at 95 °C. PCR stage (40 cycles): 15 s at 95 °C, 30 s at 58 °C, 30 s at 60 °C. Melt curve stage: 15 s at 95 °C, 1 min at 50 °C then increments of 0.05 °C/s until 95 °C. Melt curve analysis was performed to verify product homogeneity. All reactions included six biological and three technical replicates for each sample. A relative quantification method of analysis with normalization against the endogenous control rrsA and employing the primer specific efficiencies was used according to the Pfaffl method (add reference). The primers used were designed with PrimerQuest (idt). The used primer sequences were: psuK – TGCGTTAGCAGCGATTGA, AATTTACGCCTGGTGGAGTAG; arcA – GATTCATGGTACGGGACAGTAG, CCGTGACAACGAAGTCGATAA; yjtD – CGCACATGGATCTGGTGATA, GGCGTGGCGTAGTAATGATA and rrsR – GTCAGCTCGTGTTGTGAAATG, CCCACCTTCCTCCAGTTTATC.Statistics and reproducibilityCorrelation between microbiota diversity measures and E. coli loads (CFU) or persistence (1-presence or 0-absence) was performed in R using the statistical package rmcorr (version 0.5.2)56 and lme4 (version 1.1-10)57, respectively. The rate of accumulation of new ISs in vivo was compared using Wilcoxon paired signed ranked test for expected and observed insertions, while the rate of selective sweeps correlation was performed using the Spearman Correlation test. Selective sweeps were taken to be mutations or HGT events that reached  > 95% frequency in the population and kept high frequency until the end of the colonization. Statistical analysis of prophage induction as well as biofilm levels was performed using the Mann-Whitney test in GraphPad Prism (version 8.4.3). A single sample T-Test was used test if the growth rate of evolved invader clones deviates from the mean of the ancestral. A Wilcoxon rank sum test with continuity correction was used to compare the relative expression levels of the evolved clones with the ancestral. P values of x0 are of order of their inverse selection coefficient (up to logarithmic corrections):$${{{{{rm{G}}}}}}left(xright)approx 1,{{{{{rm{T}}}}}}left(xright)sim frac{1}{s}.$$
(1)

Clonal interference under uniform directional selection. This mode occurs in asexual populations when adaptive mutations become frequent enough to interfere with one another59,60,61. Only a fraction of the established adaptive mutations reaches fixation; sojourn times to intermediate frequencies are set by a global coalescence rate (widetilde{sigma }) that is higher than the typical selection coefficient of individual mutations62:

$${{{{{rm{G}}}}}}left(xright) < 1,{{{{{rm{T}}}}}}left(xright)sim frac{1}{widetilde{sigma}}.$$ (2) Details of these dynamics depend on the spectrum of selection coefficients and on the overall mutation rate, which set the strength of clonal interference. For moderate interference, where a few concurrent beneficial mutations compete for fixation, we expect a roughly exponential drop of the frequency propagator, (Gleft(xright)sim {{exp }}left(-lambda xright)), reflecting the probability that a trajectory reaches frequency x without interference by a stronger competing clade. Moderate interference generates an effective neutrality for weaker beneficial mutations and at higher frequencies63. This regime has been mapped for influenza64. In the asymptotic regime of a travelling fitness wave, where many beneficial mutations are simultaneously present, the fate of a mutation is settled in the range of small frequencies; that is, at the tip of the wave65. In this regime, emergent neutrality affects the vast majority of beneficial mutations and most of the frequency regime66. Hence, the frequency propagator rapidly drops to its asymptotic value (Gleft(x=1right)ll 1.) Adaptation under diversifying selection. More complex selection scenarios involve selection within and between ecotypes, i.e., subpopulations occupying distinct ecological niches67,68. An important factor generating niches and ecotypes is the differential use of food and other environmental resources. In this mode, ecotype-specific, conditionally beneficial mutations reach intermediate frequencies after a time given by their within-ecotype selection coefficients, but fixation can be slowed down or suppressed by diversifying (negative frequency-dependent) cross-ecotype selection18, $${{{{{rm{G}}}}}}left(xright)approx 1,{{{{{rm{T}}}}}}left(xright)sim frac{1}{s},left(xlesssim ,frac{1}{2}right)$$ (3) $${{{{{rm{G}}}}}}left(xright) < 1,{{{{{rm{T}}}}}}left(xright)gg frac{1}{s}left(xto 1right).$$ (4) The details depend on the details of the eco-evolutionary model (synergistic vs. antagonistic interactions, carrying capacities, amount of resource competition vs. explicitly frequency-dependent selection). In a model with directional selection within ecotypes, conditionally beneficial mutations rapidly fix within ecotypes, but lead only to finite shifts of the ecotype frequencies. In the simplest case, the resulting dynamics of ecotype frequencies is diffusive, resulting in an effectively neutral turnover of ecotypes18. Given negative frequency-dependent selection between ecotypes, fixations become even rarer and can be completely suppressed; that is, ecotypes can become stable on the time scales of observation. The separation of time and selection scales between intra- and cross-ecotype frequency changes is expected to be a robust feature of ecotype-dependent selection: sojourn of adaptive alleles to intermediate frequencies is fast, fixation is slower and rarer. In other words, ecotype-dependent selection is characterized by two regimes of coalescence times T(x).Frequency propagators and the coalescence time spectra expected under these evolutionary modes are qualitatively sketched in Supplementary Fig. 11. For periodic sweeps under directional selection (dark green, left column), G(x) depends weakly on x and T(x) is set by rapid sweeps for all x. For clonal interference under directional selection (green, center column), G(x) decreases substantially with increasing x and T(x) becomes uniformly shorter. Under negative frequency-dependent selection (brown, right column), G(x) decreases substantially with increasing x, while T(x) substantially increases for large x and diverges in case of strong frequency-dependent selection generating stable ecotypes (dashed lines). (see Supplementary Fig. 11 for the results of simulations assuming a model of direction selection or assuming a resource competition model where ecotype formation occurs31.The ({{{{{boldsymbol{p}}}}}})-({{{{{boldsymbol{tau }}}}}}) selection testThis test is based on qualitative characteristics of the functions G(x), T(x) and does not depend on details of the evolutionary process. We evaluate G(x) and T(x) for host-specific families of frequency trajectories; sojourn times are counted from an initial frequency x0=0.01. Origination times at this frequency are inferred by backward extrapolation of the first observed trajectory segment; the reported results are robust under variations of the threshold x0 and the extrapolation procedure. We then compute two summary statistics: the probability (p) that a mutation established at an intermediate frequency xm reaches near-fixation at a frequency xf,$$p=frac{{{{{{rm{G}}}}}}({x}_{f})}{{{{{{rm{G}}}}}}({x}_{m})},$$ (5) and the corresponding fraction of sojourn times,$$tau=,frac{{{{{{rm{T}}}}}}({x}_{f})}{{{{{{rm{T}}}}}}({x}_{m})}.$$ (6) Here we use xm=0.3 and xf=0.95 to limit the uncertainties of empirical trajectories at low and high frequency; however, the selection test is robust under variation of these frequencies. We find evidence for different modes of evolution: The long-term frequency trajectories of mice B2, D2 and E2 are consistent with predominantly frequency-dependent selection (Fig. 2, Fig. 4a–c). The propagator G(x) is a strongly decreasing function of x, resulting in fixation probabilities (p) 0.6, as measured by time ratios τ  > 3.

The trajectories of mice A2, G2, and I2 show a signature of recurrent selective sweeps and clonal interference under uniform directional selection (Fig. 4a–c). The propagator G(x) is a decreasing function of x, resulting in fixation probabilities (p=0.2-0.8), depending on the strength of clonal interference. Fixation times are short, giving time ratios (tau lesssim 2).

The shorter trajectory of mouse H2 signals periodic sweeps under uniform directional selection (Fig. 3, Fig. 4a–c). The origination rate of mutations is lower than in the longer trajectories, and G(x) shows a weak decrease with (p=1.) Sojourn times T(x) are short and grow uniformly with x, resulting in a time ratio τ=2.25. This pattern is expected under directional selection in the low mutation regime: (Tleft(xright)={{log }}left[x/(1-x)right]/{s}) for individual mutations with a uniform selection coefficient s, leading to τ=2.0 for xm=0.3 and xf=0.95 (this value is marked as a dashed line in Fig. 4c).

The trajectories of non-mutator lines in the long-term in vitro evolution experiment of Good et al1, evaluated over the first 7500 generations, show an overall signal of clonal interference under uniform directional selection (Fig. 4c, Supplementary Fig. 12). The frequency propagators G(x) are strongly decreasing functions of x and sojourn times T(x) grow uniformly with x. We find (p=0.2-0.8) and (tau lesssim 2), similar to the pattern in mice A2, G2, and I2.

Code for Selection testsThe code for selection tests from the mutation frequency trajectories can be found in the Supplementary Information file.Reporting summaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article. More