Cross-modal data integration and spectral optimization for enhanced individual apple tree canopy nitrogen concentration estimation using UAV remote sensing

Individual tree detection from airborne laser scanning data based on supervoxels and local convexity

Forests cover approximately 31% of the global land area and are home to 80% of the Earth's terrestrial biodiversity. Humans depend on forests for countless ecosystem services, but these ecosystems are highly vulnerable to human-induced climate change. As our climate undergoes dynamic changes, it is imperative to implement automated monitoring systems to quantify canopy growth and assess changes occurring within forest structures, especially at the level of individual trees, to determine the response of forests to climate anomalies. In this context, tree canopy detection can be considered one of the most important applications using Unmanned Aerial Vehicles (UAVs) as it can be used to obtain information on numerous essential ecosystem variables (EEVs) such as gross primary productivity, leaf area index, etc. for individual trees or shed light on essential biodiversity variables (EBVs) such as ecosystem structure and function. However, due to the plethora of information available, users may find it challenging to apply UAVs and algorithms to their specific projects. Hence, an integrated, seamless platform that can process UAV-acquired images to generate ortho-mosaics, detect individual trees, and monitor specific traits (including ecosystem structure and function) is the need of the hour.In this study, a platform, Drone4Tree, has been developed using Streamlit and Flask to provide an end-to-end solution for generating orthomosaics and delineating individual tree crowns from UAV images. Users simply upload raw UAV survey data and receive the final results. The complete processing chain is carried out on our high-end servers, which is an advantage for users with limited computing resources. The developed web application uses open-source algorithms, models, and frameworks for easy implementation of components such as orthomosaic (structure from motion in OpenDroneMap), tree canopy detection (DeepForest and U-Net segmentation), and downloading of results. The platform offers two processing modes: standard and advanced. The standard mode comes with default parameters for orthomosaic generation and tree canopy detection, benefiting users with no experience in UAV image processing. The advanced mode allows users to customize the processes, such as the scale of the generated canopy boundary or patch size for large images. It also extends its functionality towards analysis-ready drone image time series (incl. a co-registration of orthomosaics to a reference image using the AROSICS method and reprojection using the geospatial data abstraction library (GDAL)). Finally, the processing outcomes can be easily downloaded using the generated links. The web app was used to generate a time series of individual tree canopies, which provided a deeper understanding of changes in EEVs during a phenological cycle. The canopy boundaries can also be used to generate spectral libraries for tree species from high spatial resolution hyperspectral images, which has several applications in species detection and mapping. This platform can guide other users wishing to efficiently produce individual tree canopy boundaries for large areas without investing substantial time tailoring imagery acquisition and processing parameters. The resulting tree canopy boundaries can provide opportunities to characterize individual trees' species, size, condition, and location and are critical resources for advancing ecological theory and informing forest management.

Canopy cover is a key forest structural parameter that is commonly used in forest inventory, sustainable forest management and maintaining ecosystem services. Recently, much attention has been paid to the use of unmanned aerial vehicle (UAV)-based light detection and ranging (LiDAR) due to the flexibility, convenience, and high point density advantages of this method. In this study, we used UAV-based LiDAR data with individual tree segmentation-based method (ITSM), canopy height model-based method (CHMM), and a statistical model method (SMM) with LiDAR metrics to estimate the canopy cover of a pure ginkgo (Ginkgo biloba L.) planted forest in China. First, each individual tree within the plot was segmented using watershed, polynomial fitting, individual tree crown segmentation (ITCS) and point cloud segmentation (PCS) algorithms, and the canopy cover was calculated using the segmented individual tree crown (ITSM). Second, the CHM-based method, which was based on the CHM height threshold, was used to estimate the canopy cover in each plot. Third, the canopy cover was estimated using the multiple linear regression (MLR) model and assessed by leave-one-out cross validation. Finally, the performance of three canopy cover estimation methods was evaluated and compared by the canopy cover from the field data. The results demonstrated that, the PCS algorithm had the highest accuracy (F = 0.83), followed by the ITCS (F = 0.82) and watershed (F = 0.79) algorithms; the polynomial fitting algorithm had the lowest accuracy (F = 0.77). In the sensitivity analysis, the three CHM-based algorithms (i.e., watershed, polynomial fitting and ITCS) had the highest accuracy when the CHM resolution was 0.5 m, and the PCS algorithm had the highest accuracy when the distance threshold was 2 m. In addition, the ITSM had the highest accuracy in estimation of canopy cover (R2 = 0.92, rRMSE = 3.5%), followed by the CHMM (R2 = 0.94, rRMSE = 5.4%), and the SMM had a relative low accuracy (R2 = 0.80, rRMSE = 5.9%).The UAV-based LiDAR data can be effectively used in individual tree crown segmentation and canopy cover estimation at plot-level, and CC estimation methods can provide references for forest inventory, sustainable management and ecosystem assessment.

The influence of canopy composition on litterfall and throughfall was investigated in a mixed spruce beech forest in central Germany. We hypothesised that different parts of the mixed canopy created distinct patterns of element inputs via litterfall and throughfall. The investigation was carried out in two plots, representing the most contrasting cases of mixed forests: a stand greatly dominated by spruce (SDP) and a stand greatly dominated by beech (BDP). The canopies of the two plots were classified in four categories: pure beech, pure spruce, mixed canopy and gap. Amounts of throughfall water were lower and major element fluxes were higher under spruce than under beech in both plots, indicating that the nutrient inputs under the canopies of individual trees are driven by species-specific properties of the canopies and are quite independent of the degree of admixture. With the exception of K+, mixed canopies showed intermediate element inputs via throughfall, compared with pure canopy classes. The K+ input was significantly greater under mixed canopies, and these differences were more pronounced in the SDP than in the BDP. Results suggest that individual spruce trees in the BDP induce greater spatial heterogeneity of throughfall input than individual beech trees in the SDP. Nutrient inputs via foliar litterfall were similar among the different canopy classes, but the Mg input was lower under spruce canopy. This effect was balanced by higher Mg input via spruce throughfall. In our study, throughfall was the main source of heterogeneity in nutrient inputs, while foliar litterfall had a homogenising effect.

It was hypothesised that in mixed spruce-beech forest ecosystems the heterogeneity in canopy composition may create different micro-sites (units) in the forest floor and the mineral soil with different ecological characteristics. Therefore, different types of canopy compositions (canopy classes) were used to identify the variability of water and element fluxes (via throughfall and litterfall), soil and soil solution chemistry, litter decomposition and soil respiration. The investigation was carried out in a mixed sprucebeech stand in Solling, central Germany. Two different plots were selected for this study representing the most contrasting cases of mixed forests types, which were i) a site greatly dominated by spruce trees (the spruce dominated plot, SDP) with two beech trees in-between and ii) a directly neighbouring site which was dominated by beech trees and having a single spruce tree in-between (the beech dominated plot, BDP). The canopies of the two plots were classified in four categories: pure beech, pure spruce, mixed canopy and gap. Throughfall water was significantly lower and major element fluxes were higher under spruce than under beech in both plots. This indicated that the nutrient inputs under the canopies of individual trees were driven by species-specific properties of the canopies and were quite independent of the degree of admixture. With the exception of K+, mixed canopies showed intermediate element inputs via throughfall, compared with pure canopy classes. The K+ input, however, was significantly greater under mixed canopies due to interactions of the canopies, leading to higher leaching rates for K+. Throughfall was the main source of heterogeneity in nutrient inputs, while foliar litterfall input was almost equal between sub-plots and thus had a homogenising effect on annual nutrient fluxes in the beech-spruce mixed stands. Differences in soil chemistry under different canopy classes were mainly observed in the forest floor and top mineral soil layers. Significant effects of the canopy composition on pH (CaCl2) values of the forest floor and mineral soil were detected between the gap (significantly higher) and spruce (significantly lower) sub-plots in the spruce dominated plot (SDP). The water fluxes (lower under spruce) and chemistry (higher concentration of elements under spruce) of throughfall could explain theses differences. In spite of almost equal litterfall inputs, different masses of organic matter (humuslayer) were observed in the forest floor of different sub-plots for both plots, SDP and BDP. Differences were most pronounced between the spruce, beech and gap sub-plots (spruce ≥ beech ≥ gap). The soil solution at 10 cm soil depth showed significantly higher pH values in the beech sub-plots, compared with the spruce sub-plots. This finding may be linked to different water and element fluxes via throughfall between sub-plots. A significant effect of the canopy composition on the rate of litter decay and the soil CO2 efflux was observed in the beech dominated plot (BDP). Here, the beech and gap sub-plots showed significantly lower remaining masses at the end of the incubation period (about one year after incubation) compared with the spruce sub-plot. This may indicate that the early stage of the decomposition process was not governed by the given canopy composition. The beech sub-plot showed significantly higher soil respiration, compared with the gap sub-plot. An estimation of the root-associated CO2 production revealed considerably lower root respiration in the gap sub-plot compared with the other sub-plots in the BDP. In total, it was shown that the selected canopy classes were able to create specific biogeochemical patterns in the investigated mixed beech-spruce forest. However, the impact of an individual spruce tree in a beech-dominated site induced obviously higher degrees of spatial heterogeneity with respect to nutrient inputs via throughfall, litter decomposition and soil respiration compared to individual beech trees in a spruce dominated site.

Lidar (light detection and ranging) has been widely utilized for estimating the structural parameters of plants, such as tree height, leaf inclination angle, and biomass. However, individual trees have been primarily manually extracted from three-dimensional (3D) point cloud images. Automatically detecting each tree and analyzing its structural parameters is desirable. In this study, we propose a method to (1) detect each tree from 3D point cloud images obtained from ground-based lidar, (2) estimate the number of trees and diameter at breast height (DBH) from the detected 3D point cloud images of trees, and (3) segment each tree canopy. First, we focused on point clouds whose height ranged from 0.5 to 1.5 m and detected each cluster of tree trunks. Then, the clusters were expanded by classifying other points to the clusters that are located near the points and then repeating this process. The process assigns the points in the 3D point cloud image to each tree in the upward direction and separates not only tree trunks but also tree canopies. As a result, the trees in 3D point cloud images were detected with high accuracy, and the number of trees and DBH was estimated. Moreover, each tree canopy was segmented.

Accurate monitoring and spatial distribution of the leaf chlorophyll content (LCC) and canopy chlorophyll content (CCC) of individual apple trees are highly important for the effective management of individual plants and the promotion of the construction of modern smart orchards. However, the estimation of LCC and CCC is affected by shadows caused by canopy structure and observation geometry. In this study, we resolved the response relationship between individual apple tree crown spectra and shadows through a three-dimensional radiative transfer model (3D RTM) and unmanned aerial vehicle (UAV) multispectral images, assessed the resistance of a series of vegetation indices (VIs) to shadows and developed a hybrid inversion model that is resistant to shadow interference. The results revealed that (1) the proportion of individual tree canopy shadows exhibited a parabolic trend with time, with a minimum occurring at noon. Correspondingly, the reflectance in the visible band decreased with increasing canopy shadow ratio and reached a maximum value at noon, whereas the pattern of change in the reflectance in the near-infrared band was opposite that in the visible band. (2) The accuracy of chlorophyll content estimation varies among different VIs at different canopy shadow ratios. The top five VIs that are most resistant to changes in canopy shadow ratios are the NDVI-RE, Cire, Cigreen, TVI, and GNDVI. (3) For the constructed 3D RTM ​+ ​GPR hybrid inversion model, only four VIs, namely, NDVI-RE, Cire, Cigreen, and TVI, need to be input to achieve the best inversion accuracy. (4) Both the LCC and the CCC of individual trees had good validation accuracy (LCC: R2 ​= ​0.775, RMSE ​= ​6.86 ​μg/cm2, nRMSE ​= ​12.24 ​%; CCC: R2 ​= ​0.784, RMSE ​= ​32.33 ​μg/cm2, and nRMSE ​= ​14.49 ​%), and their distributions at orchard scales were characterized by considerable spatial heterogeneity. This study provides ideas for investigating the response between individual tree canopy shadows and spectra and offers a new strategy for minimizing the influence of shadow effects on the accurate estimation of chlorophyll content in individual apple trees.

In this study, we confirmed the utility of airborne and portable on-ground scanning light detection and ranging (LIDARs) for three-dimensional visualization of an urban park and quantification of biophysical variables of trees in the park. The digital canopy height model (DCHM) and digital terrain model generated from airborne scanning LIDAR data provided precise images of the ground surface and individual tree canopies. The heights of 166 coniferous and broadleaf trees of 11 species in the park were estimated from the DCHM images with slight underestimation (mean error = −0.14 m, RMSE = 0.30 m). Portable on-ground scanning LIDAR provided images of individual trees with detailed features. Tree height and trunk diameter were estimated to be within 0.31 m and 1 cm, respectively, from the on-ground LIDAR images. We combined airborne and on-ground LIDAR images to overcome blind regions and created a complete three-dimensional model of three standing trees. The model allowed not only visual assessment from all viewpoints but also quantitative estimation of canopy volume, trunk volume, and canopy cross-sectional area.

It is highly necessary to apply unmanned aerial vehicle (UAV) remote sensing technology to forest health assessment. To prove the feasibility of quantitative inversion of photosynthetic pigment content (PPC) in Populus euphratica Oliv. individual tree canopy (PeITC) by using multispectral UAV images, in this study, Parrot Sequoia+ multispectral UAV system was manipulated to collect the images of Populus euphratica (Populus euphratica Oliv.) sample plots in Daliyabuyi Oasis from 2019 to 2020, and the canopy PPCs of five Populus euphratica sample trees per plot were determined in six plots. The Populus euphratica crown regions were extracted by grey wolf optimizer-OTSU (GWO-OTSU) multithreshold segmentation algorithm from the normalized difference vegetation index (NDVI) images of Populus euphratica sample plots obtained after preprocessing, and the PeITCs were segmented by multiresolution segmentation algorithm. The mean values of 27 spectral indices in the PeITCs were calculated in each plot, and the optimal model was constructed for quantitative estimation of the PPCs in the PeITCs, then the inversion results were compared and verified based on GF-6 and ZY1-02D satellite imageries respectively. The results were as follows. (1) The average value of canopy chlorophyll content (Chl) was 2.007 mg/g, the mean value of canopy carotenoid content (Car) was 0.703 mg/g. The coefficient of variation (C.V) of both were basically the same and they were both of strong variability. The measured PPCs of the PeITCs in Daliyabuyi Oasis was generally low. The average contents of chlorophyll and carotenoid in PeITC in June were more than twice those in August, while the mean ratio between them was significantly lower in June than in August. The measured PPCs had no obvious spatial distribution law. However, that could prove the rationality of sample selection in this study. (2) NDVI had the best effect of highlighting vegetation among all quadrats in the study area. Based on the GWO-OTSU multithreshold segmentation method, the canopy area of Populus euphratica could be quickly and effectively extracted from the quadrat NDVI map. The best segmentation effect of PeITCs was obtained based on a multiresolution segmentation method when the segmentation scale was 120, the shape index was 0.7, and the compactness index was 0.5. Compared with manual vectorization method of visual interpretation, the root mean square error (RMSE) and Pearson correlation coefficient (R) values of the mean NDVI values in PeITCs obtained by these two methods were 0.038 and 0.951. (3) Only 12 of the 27 spectral indices were significantly correlated with Chl and Car at the significance level of 0.02. Characteristics of the calibration set and validation set were basically consistent with those of the entire set. The classification and regression tree-decision tree (CART-DT) model performed best in the estimation of the PPCs in the PeITCs, in which, when estimating the Car, the calibration coefficient of determination (R2C) was 0.843, the calibration root mean square error (RMSEC) was 0.084, the calibration residual prediction deviation (RPDC) was 2.525, the validation coefficient of determination (R2V) was 0.670, the validation root mean square error (RMSEV) was 0.251, the validation residual prediction deviation (RPDV) was 1.741. (4) Qualitative comparison of spectral reflectance and NDVI values between GF-6 multispectral imagery and Parrot Sequoia+ multispectral image on the 172 PeITCs can show the reliability of Parrot Sequoia+ multispectral image. The comparison results of five PeITCs relative health degree judged by field vision judgment, measured SPAD value, predicted value of Chl (Chlpre), the red edge value calculated by ZY1-02D (ZY1-02Dred edge) and the Carotenoid Reflection Index 2 (CRI2) value calculated by ZY1-02D (ZY1-02DCRI2) can further prove the scientificity of inversion results to a certain extent. These results indicate that multispectral UAV images can be applied for quantitative inversion of PPC in PeITC, which could provide an indicator for the construction of a Populus euphratica individual tree health evaluation indicator system based on UAV remote sensing technology in the next step.

Automatic 3D forest mapping and individual tree characteristics estimation are essential for forest management and ecosystem maintenance. The low-cost unmanned aerial vehicle (UAV) laser scanning (ULS) is a newly developed tool for cost-effectively collecting 3D information and attempts to use it for 3D forest mapping have been made, due to its capability to provide 3D information with a lower cost and higher flexibility than the standard ULS and airborne laser scanning (ALS). As the direct georeferenced point clouds may suffer from distortion caused by the poor performance of a low-cost inertial measurement unit (IMU), and 3D forest mapping using low-cost ULS poses a great challenge. Therefore, this paper utilized global navigation satellite system (GNSS) and IMU aided Structure-from-Motion (SfM) for trajectory estimation, and, hence, overcomes the poor performance of low-cost IMUs. The accuracy of the low-cost ULS point clouds was compared with the ground truth data collected by a commercial ULS system. Furthermore, the effectiveness of individual trees segmentation and tree characteristics estimation derived from the low-cost ULS point clouds were accessed. Experiments were undertaken in Dongtai forest farm, Yancheng City, Jiangsu Province, China. The results showed that the low-cost ULS achieved good point clouds quality from visual inspection and comparable individual tree segmentation results (P = 0.87, r = 0.84, F = 0.85) with the commercial system. Individual tree height estimation performed well (coefficient of determination (R2) = 0.998, root-mean-square error (RMSE) = 0.323 m) using the low-cost ULS. As for individual tree crown diameter estimation, low-cost ULS achieved good results (R2 = 0.806, RMSE = 0.195 m) after eliminating outliers. In general, such results illustrated the high potential of the low-cost ULS in 3D forest mapping, even though 3D forest mapping using the low-cost ULS requires further research.

Intra- and interplant leaf sesquiterpene variability in Copaifera langsdorfii: relation to microlepidopteran herbivory

Accurate forest parameters are crucial for ecological protection, forest resource management and sustainable development. The rapid development of remote sensing can retrieve parameters such as the leaf area index, cluster index, diameter at breast height (DBH) and tree height at different scales (e.g., plots and stands). Although some LiDAR satellites such as GEDI and ICESAT-2 can measure the average tree height in a certain area, there is still a lack of effective means for obtaining individual tree parameters using high-resolution satellite data, especially DBH. The objective of this study is to explore the capability of 2D image-based features (texture and spectrum) in estimating the DBH of individual tree. Firstly, we acquired unmanned aerial vehicle (UAV) LiDAR point cloud data and UAV RGB imagery, from which digital aerial photography (DAP) point cloud data were generated using the structure-from-motion (SfM) method. Next, we performed individual tree segmentation and extracted the individual tree crown boundaries using the DAP and LiDAR point cloud data, respectively. Subsequently, the eight 2D image-based textural and spectral metrics and 3D point-cloud-based metrics (tree height and crown diameters) were extracted from the tree crown boundaries of each tree. Then, the correlation coefficients between each metric and the reference DBH were calculated. Finally, the capabilities of these metrics and different models, including multiple linear regression (MLR), random forest (RF) and support vector machine (SVM), in the DBH estimation were quantitatively evaluated and compared. The results showed that: (1) The 2D image-based textural metrics had the strongest correlation with the DBH. Among them, the highest correlation coefficient of −0.582 was observed between dissimilarity, variance and DBH. When using textural metrics alone, the estimated DBH accuracy was the highest, with a RMSE of only 0.032 and RMSE% of 16.879% using the MLR model; (2) Simply feeding multi-features, such as textural, spectral and structural metrics, into the machine learning models could not have led to optimal results in individual tree DBH estimations; on the contrary, it could even reduce the accuracy. In general, this study indicated that the 2D image-based textural metrics have great potential in individual tree DBH estimations, which could help improve the capability to efficiently and meticulously monitor and manage forests on a large scale.

Abstract. Vertical profiles in leaf mass per unit leaf area (MA), foliar 13C composition (δ13C), nitrogen (N), phosphorus (P), carbon (C) and major cation concentrations were estimated for 204 rain forest trees growing in 57 sites across the Amazon Basin. Data was analysed using a multilevel modelling approach, allowing a separation of gradients within individual tree canopies (within-tree gradients) as opposed to stand level gradients occurring because of systematic differences occurring between different trees of different heights (between-tree gradients). Significant positive within-tree gradients (i.e. increasing values with increasing sampling height) were observed for MA and [C]DW (the subscript denoting on a dry weight basis) with negative within-tree gradients observed for δ13C, [Mg]DW and [K]DW. No significant within-tree gradients were observed for [N]DW, [P]DW or [Ca]DW. The magnitudes of between-tree gradients were not significantly different to the within-tree gradients for MA, δ13C, [C]DW, [K]DW, [N]DW, [P]DW and [Ca]DW. But for [Mg]DW, although there was no systematic difference observed between trees of different heights, strongly negative within-tree gradients were found to occur. When expressed on a leaf area basis (denoted by the subscript "A"), significant positive gradients were observed for [N]A, [P]A and [K]A both within and between trees, these being attributable to the positive intra- and between-tree gradients in MA mentioned above. No systematic within-tree gradient was observed for either [Ca]A or [Mg]A, but with a significant positive gradient observed for [Mg]A between trees (i.e. with taller trees tending to have a higher Mg per unit leaf area). Significant differences in within-tree gradients between individuals were observed only for MA, δ13C and [P] A. This was best associated with the overall average [P]A for each tree, this also being considered to be a surrogate for a tree's average leaf area based photosynthetic capacity, Amax. A new model is presented which is in agreement with the above observations. The model predicts that trees characterised by a low upper canopy Amax should have shallow, or even non-existent, within-canopy gradients in Amax, with optimal intra-canopy gradients becoming sharper as a tree's upper canopy Amax increases. Nevertheless, in all cases it is predicted that the optimal within-canopy gradient in Amax should be substantially less than for photon irradiance. Although this is also shown to be consistent with numerous observations as illustrated by a literature survey of gradients in photosynthetic capacity for broadleaf trees, it is also in contrast to previously held notions of optimality. A new equation relating gradients in photosynthetic capacity within broadleaf tree canopies to the photosynthetic capacity of their upper canopy leaves is presented.

To reduce data acquisition cost, this study proposed a novel method of individual tree height estimation and canopy extraction based on fusion of an airborne multispectral image and photogrammetric point cloud. A fixed‐wing drone was deployed to acquire the true color and multispectral images of a shelter forest. The Structure‐from‐Motion (SfM) algorithm was used to reconstruct the 3D point cloud of the canopy. The 3D point cloud was filtered to acquire the ground point cloud and then interpolated to a Digital Elevation Model (DEM) using the Radial Basis Function Neural Network (RBFNN). The DEM was subtracted from the Digital Surface Model (DSM) generated from the original point cloud to get the canopy height model (CHM). The CHM was processed for the crown extraction using local maximum filters and watershed segmentation. Then, object‐oriented methods were employed in the combination of 12 bands and CHM for image segmentation. To extract the tree crown, the Support Vector Machine (SVM) algorithm was used. The result of the object‐oriented method was vectorized and superimposed on the CHM to estimate the tree height. Experimental results demonstrated that it is efficient to employ point cloud and the proposed approach has great potential in the tree height estimation. The proposed object‐oriented method based on fusion of a multispectral image and CHM effectively reduced the oversegmentation and undersegmentation, with an increase in the F‐score by 0.12–0.17. Our findings provided a reference for the health and change monitoring of shelter forests as well.

The ecology of forest ecosystems depends on the composition of trees. Capturing fine-grained information on individual trees at broad scales provides a unique perspective on forest ecosystems, forest restoration, and responses to disturbance. Individual tree data at wide extents promises to increase the scale of forest analysis, biogeographic research, and ecosystem monitoring without losing details on individual species composition and abundance. Computer vision using deep neural networks can convert raw sensor data into predictions of individual canopy tree species through labeled data collected by field researchers. Using over 40,000 individual tree stems as training data, we create landscape-level species predictions for over 100 million individual trees across 24 sites in the National Ecological Observatory Network (NEON). Using hierarchical multi-temporal models fine-tuned for each geographic area, we produce open-source data available as 1 km2 shapefiles with individual tree species prediction, as well as crown location, crown area, and height of 81 canopy tree species. Site-specific models had an average performance of 79% accuracy covering an average of 6 species per site, ranging from 3 to 15 species per site. All predictions are openly archived and have been uploaded to Google Earth Engine to benefit the ecology community and overlay with other remote sensing assets. We outline the potential utility and limitations of these data in ecology and computer vision research, as well as strategies for improving predictions using targeted data sampling.