Remote Sensing in Ecology and Conservation

Tree detection is a fundamental task in remote sensing for forestry and ecosystem ecology applications. While many individual tree segmentation algorithms have been proposed, the development and testing of these algorithms is typically site specific, with few methods evaluated against data from multiple forest types simultaneously. This makes it difficult to determine the generalization of proposed approaches, and limits tree detection at broad scales. Using data from the National Ecological Observatory Network we extend a recently developed semi-supervised deep learning algorithm to include data from a range of forest types, determine whether information from one forest can be used for tree detection in other forests, and explore the potential for building a universal tree detection algorithm. We find that the deep learning approach works well for overstory tree detection across forest conditions, outperforming conventional LIDAR-only methods in all forest types. Performance was best in open oak woodlands and worst in alpine forests. When models were fit to one forest type and used to predict another, performance generally decreased, with better performance when forests were more similar in structure. However, when models were pretrained on data from other sites and then fine-tuned using a small amount of hand-labeled data from the evaluation site, they performed similarly to local site models. Most importantly, a universal model fit to data from all sites simultaneously performed as well or better than individual models trained for each local site. This result suggests that RGB tree detection models that can be applied to a wide array of forest types at broad scales should be possible.

O_LIEcological data are increasingly collected over vast geographic areas using arrays of digital sensors. Camera trap arrays have become the gold standard method for surveying many terrestrial mammals and birds, but these arrays often generate millions of images that are challenging to process. This causes significant latency between data collection and subsequent inference, which can impede conservation at a time of ecological crisis. Machine learning algorithms have been developed to improve camera trap data processing speeds, but these models are not considered accurate enough for fully automated labeling of images. C_LIO_LIHere, we present a new approach to building and testing a high performance machine learning model for fully automated labeling of camera trap images. As a case-study, the model classifies 26 Central African forest mammal and bird species (or groups). The model was trained on a relatively small dataset (c.300,000 images) but generalizes to fully independent data and outperforms humans in several respects (e.g. detecting invisible animals). We show how the models precision and accuracy can be evaluated in an ecological modeling context by comparing species richness, activity patterns (n = 4 species tested) and occupancy (n = 4 species tested) derived from machine learning labels with the same estimates derived from expert labels. C_LIO_LIResults show that fully automated labels can be equivalent to expert labels when calculating species richness, activity patterns (n = 4 species tested) and estimating occupancy (n = 3 of 4 species tested) in completely out-of-sample test data (n = 227 camera stations, n = 23868 images). Simple thresholding (discarding uncertain labels) improved the models performance when calculating activity patterns and estimating occupancy, but did not improve estimates of species richness. C_LIO_LIWe provide the user-community with a multi-platform, multi-language user interface for running the model offline, and conclude that high performance machine learning models can fully automate labeling of camera trap data. C_LI

Woody plant encroachment (WPE)--a phenomenon similar to species invasion--is shifting many grasslands and savannas into shrub and evergreen-dominated ecosystems. Tracking WPE is difficult because shrubs and small trees are much smaller than the coarse resolution of common remote sensing platforms (> 10 m2) and the impassibility of encroaching woody thickets slows ground-based approaches. Many agencies have been investing in fine resolution (< 2 m2) remote sensing through programs such as the United States Department of Agriculture (USDA) National Agriculture Imagery Program (NAIP) and the National Ecological Observatory Network (NEON). Both use low-flying planes and provide data to end users in an easy-to-use format at large spatial extents. By removing entry barriers, these publicly available open-source programs could increase the accessibility and extent of remote sensing. We compared two common methods of machine learning classification of land cover (random forests and support vector machines) factorially crossed with these two freely available remote sensing platforms to determine if we could quickly and accurately develop remote sensing of major vegetation types in a tallgrass prairie landscape undergoing encroachment by shrubs and trees. Our work took place at Konza Prairie Biological Station--a landscape scale experiment that results in a wide range of land cover types. All models had very high overall classification accuracy (>90%), with the NEON-based models a few percent more accurate than NAIP. A model using both inputs had the highest accuracy. However, the accuracies of NAIP and NEON models differed for woody vegetation: compared to NEON, NAIP accuracy was, 82-93% compared to 94-98% for shrubs, 72-92% compared to 93-98% for deciduous trees, and 52-78% compared to 83-86% for evergreen trees (specifically Juniperus virginiana). NEON-based models relied on canopy height (LiDAR) to make classifications, whereas the several bands of light make similar contributions to accuracy in the NAIP models. Finally, we found that both machine learning approaches had similar accuracy, but random forests ran substantially faster. We conclude that with large training datasets, publicly available aerial imagery and similar products (e.g., UAVs, micro-satellites) can produce fine-scale, high-accuracy remote sensing of WPE in this region with low up-front costs.

Delineating and classifying individual trees in remote sensing data is challenging. Many tree crown delineation methods have difficulty in closed-canopy forests and do not leverage multiple datasets. Methods to classify individual species are often accurate for common species, but perform poorly for less common species and when applied to new sites. We ran a data science competition to help identify effective methods for delineation of individual crowns and classification to determine species identity. This competition included data from multiple sites to assess the methods ability to generalize learning across multiple sites simultaneously, and transfer learning to novel sites where the methods were not trained. Six teams, representing 4 countries and 9 individual participants, submitted predictions. Methods from a previous competition were also applied and used as the baseline to understand whether the methods are changing and improving over time. The best delineation method was based on an instance segmentation pipeline, closely followed by a Faster R-CNN pipeline, both of which outperformed the baseline method. However, the baseline (based on a growing region algorithm) still performed well as did the Faster R-CNN. All delineation methods generalized well and transferred to novel forests effectively. The best species classification method was based on a two-stage fully connected neural network, which significantly outperformed the baseline (a random forest and Gradient boosting ensemble). The classification methods generalized well, with all teams training their models using multiple sites simultaneously, but the predictions from these trained models generally failed to transfer effectively to a novel site. Classification performance was strongly influenced by the number of field-based species IDs available for training the models, with most methods predicting common species well at the training sites. Classification errors (i.e., species misidentification) were most common between similar species in the same genus and different species that occur in the same habitat. The best methods handled class imbalance well and learned unique spectral features even with limited data. Most methods performed better than baseline in detecting new (untrained) species, especially in the site with no training data. Our experience further shows that data science competitions are useful for comparing different methods through the use of a standardized dataset and set of evaluation criteria, which highlights promising approaches and common challenges, and therefore advances the ecological and remote sensing field as a whole.

Accurate, up-to-date catalogs of urban tree populations are crucial for quantifying ecosystem services and enhancing the quality of life in cities. However, mapping tree species cost-effectively remains challenging. In response, remote sensing researchers are developing general-purpose tools to survey plant populations across broad spatial scales. In this study, we developed computer vision models to detect, classify, and map 100 tree genera across 23 cities in North America using Google Street View (GSV) and iNaturalist images. We validated our predictions in independent portions of each city. We then compared our predictions to existing street tree records to evaluate the spatial context of errors using generalized linear mixed-effects models. Our computer vision models identified most ground-truthed street trees (67.1%). Performance varied across the 23 cities (67.4% {+/-} 9.3%) and 100 genera (50.9% {+/-} 23.0%) and improved denser street-view coverage, simpler stand structure, and greater training representation, particularly from the focal city. We found that genus classification performed better in continental cities with lower relative diversity, and that seasonal changes in the appearance of trees provided visual cues that moderate classification rates. Using widely available street-level imagery is a generalizable and promising avenue for mapping tree distributions across urban environments.

The Tasmanian masked owl (Tyto novaehollandiae subsp. castanops) and swift parrot (Lathamus discolor), both endangered species, rely on old forest features that are declining across their ranges in Tasmania, Australia. Their elusive behavior and high mobility make monitoring difficult, hindering conservation actions. Passive acoustic monitoring can greatly increase spatial and temporal survey coverage, though the identification of the species vocalisations within large audio datasets remains challenging. We deployed automated recording units at 108 sites in Tasmanias native forests to collect a large and representative acoustic dataset. We trained a neural network model to automate call detection for both the Tasmanian masked owl and the swift parrot. Our model demonstrated high performance, with a sensitivity of 97.7% and specificity of 96.5% for the Tasmanian masked owl, and a sensitivity of 87.5% and specificity of 83.3% for the swift parrot. Through two real-world applications, we illustrated how our method provides detailed quantitative insights into habitat use patterns over extended spatial and temporal scales. This innovative approach enhances both site-specific and population monitoring, enabling more effective and targeted conservation actions for these endangered species.

O_LIThe ability to measure flying insect activity and abundance is important for ecologists, conservationists and agronomists alike. However, existing methods are laborious and produce data with low temporal resolution (e.g. trapping and direct observation), or are expensive, technically complex, and require vehicle access to field sites (e.g. radar and lidar entomology). C_LIO_LIWe propose a method called "camfi" for long-term non-invasive monitoring of the activity and abundance of low-flying insects using images obtained from inexpensive wildlife cameras, which retail for under USD$100 and are simple to operate. We show that in certain circumstances, this method facilitates measurement of wingbeat frequency, a diagnostic parameter for species identification. To increase usefulness of our method for very large monitoring programs, we have developed and implemented a tool for automatic detection and annotation of flying insect targets based on the popular Mask R-CNN framework. This tool can be trained to detect and annotate insects in a few hours, taking advantage of transfer learning. C_LIO_LIWe demonstrate the utility of the method by measuring activity levels and wingbeat frequencies in Australian Bogong moths Agrotis infusa in the Snowy Mountains of New South Wales, and find that these moths have log-normally distributed wingbeat frequencies (mean = 49.4 Hz, std = 5.25 Hz), undertake dusk flights in large numbers, and that the intensity of their dusk flights is modulated by daily weather factors. Validation of our tool for automatic image annotation gives baseline performance metrics for comparisons with future annotation models. The tool performs well on our test set, and produces annotations which can be easily modified by hand if required. Training completed in less than 2 h on a single machine, and inference took on average 1.15 s per image on a laptop. C_LIO_LIOur method will prove invaluable for ongoing efforts to understand the behaviour and ecology of the iconic Bogong moth, and can easily be adapted to other flying insects. The method is particularly suited to studies on low-flying insects in remote areas, and is suitable for very large-scale monitoring programs, or programs with relatively low budgets. C_LI

O_LICamera trap arrays can generate thousands to millions of images that require exorbitant time and effort to classify and annotate by trained observers. Computer vision has evolved as an automated alternative to manual classification. The most popular computer vision solution is the supervised Machine Learning technique, which uses labeled images to train automated classification algorithms. C_LIO_LIWe propose a multi-step semi-automated workflow that consists of (1) identifying and separating bad-from good-quality images, (2) parsing good images into animals, humans, vehicles, and empty, and (3) cropping animals from images and classifying them into species for manual inspection. We trained, validated, and evaluated this approach using 548,627 images from 46 cameras in two regions of the Arctic (northeastern Norway, and Yamal Peninsula, Russia). C_LIO_LIWe obtained an accuracy of 0.959 for all three steps combined with the complete year test data set at Varanger and 0.922 at Yamal, reducing the number of images that required manual inspection to 7.9% of the original set from Varanger and 3.2% from Yamal. C_LIO_LIResearchers can modify this multi-step process to meet their specific needs for monitoring and surveying wildlife, providing greater flexibility than current options available for image classification. C_LI

With bird populations across the world being impacted by ever-growing anthropogenic pressures, reliable monitoring is essential to help halt or reverse declines. Existing visual bird monitoring approaches, which employ cameras or radars to deliver automated and large-scale monitoring data, face a variety of issues. Image-based species classification is only possible if the fine-scale features of a bird are clear, which can be difficult to achieve in real monitoring contexts without expensive, high-resolution cameras due to occlusion and lighting. Radar and video-based approaches which analyse longer-term flight behaviour over the course of seconds can achieve more reliable results in real monitoring contexts, particularly from greater distances, but still require expensive equipment and do not account for all the possible types of flight patterns. Here we present a novel approach to track a wide range of bird flight patterns using inexpensive equipment. As a proof-of-concept, we demonstrate how our approach can be used to classify birds between four species, Red Kite, Kestrel, Black-Headed Gull and Sparrowhawk, which represent four different types of flight patterns. The balanced accuracy of the classification is 0.5583, with a recall and precision per species that range from 0.2640-0.7750 and 0.4583-0.5962, respectively. Our proof-of-concept study demonstrates how new and existing visual bird monitoring systems can leverage flight patterns to deliver species-level insights at lower costs and on larger scales than before.

Accurate monitoring of populations is essential for conservation management, including for vulnerable seabirds. Yet traditional ground-based surveys are logistically challenging and time-consuming, especially in remote environments such as the sub-Antarctic islands. Advances in aerial imagery and artificial intelligence (AI) offer opportunities to improve the efficiency and repeatability of population surveys. In this study, we evaluate an AI-based approach for counting Salvins albatross from high-resolution aerial imagery collected using a piloted fixed-wing aircraft at the Bounty Islands, New Zealand. Imagery acquired during a single-day survey was processed to create orthomosaic images, which were previously analysed using manual counts by an experienced observer. We applied an automated detection and counting model based on a Faster R-CNN architecture with Slicing-Aided Hyper-Inference, and compared AI-derived counts with original human counts in terms of accuracy, consistency, and processing time. The AI achieved an initial F1 score of 92.8% for albatross detection and produced counts within 3% of the manual results, while reducing processing time from approximately 66 hours to just over four minutes. The model was also capable of simultaneously detecting additional species present within the mixed breeding colony, including erect-crested penguins, fulmar prions, and New Zealand fur seals, adding scalable efficiency gains for future surveys. Our results demonstrate that combining piloted aircraft surveys with AI-based image analysis provides a rapid, scalable, and accurate method for monitoring seabird populations, with substantial benefits for conservation management in remote and logistically constrained regions.

Image-based machine learning methods are quickly becoming among the most widely-used forms of data analysis across science, technology, and engineering. These methods are powerful because they can rapidly and automatically extract rich contextual and spatial information from images, a process that has historically required a large amount of manual labor. The potential of image-based machine learning methods to change how researchers study the ocean has been demonstrated through a diverse range of recent applications. However, despite their promise, machine learning tools are still under-exploited in many domains including species and environmental monitoring, biodiversity surveys, fisheries abundance and size estimation, rare event and species detection, the study of wild animal behavior, and citizen science. Our objective in this article is to provide an approachable, application-oriented guide to help researchers apply image-based machine learning methods effectively to their own research problems. Using a case study, we describe how to prepare data, train and deploy models, and avoid common pitfalls that can cause models to underperform. Importantly, we discuss how to diagnose problems that can cause poor model performance on new imagery to build robust tools that can vastly accelerate data acquisition in the marine realm. Code to perform our analyses is provided at https://github.com/heinsense2/AIO_CaseStudy

Reported insect declines have dramatically increased the global demand for standardized insect monitoring data. Image-based monitoring can generate such data cost-efficiently and non-invasively. However, extracting ecological data from images is more challenging for insects than for vertebrates because of their small size and great diversity. Deep learning facilitates fast and accurate insect detection and identification, but the lack of training data for coveted deep learning models is a major obstacle for their application. We present a large annotated image dataset of functionally important insect taxa. The primary dataset consists of 29,960 annotated insects representing nine taxa including bees, hoverflies, butterflies and beetles across more than two million images recorded with ten time-lapse cameras mounted over flowers during the summer of 2019. The insect image dataset was extracted using an iterative approach: First, a preliminary detection model identified candidate insects. Second, candidate insects were manually screened by users of an online citizen science platform. Finally, all annotations were quality checked by experts. We used the dataset to train and compare the performance of selected You Only Look Once (YOLO) deep learning algorithms. We show that these models detect and classify small insects in complex scenes with unprecedented accuracy. The best performing YOLOv5 model consistently identifies nine dominant insect species that play important roles in pollination and pest control across Europe. The model reached an average precision of 92.7% and recall of 93.8 % in detection and classification across species. Importantly, when presented with uncommon or unclear insects not seen during training, our model detects 80% of individuals and usually interprets them as closely related species. This is a useful property to (1) detect rare insects for which training data are absent, and (2) generate new training data to correctly identify those insects in future. Our camera system, dataset and deep learning framework show promising results in non-destructive monitoring of insects. Furthermore, resulting data are useful to quantify phenology, abundance, and foraging behaviour of flower-visiting insects. Above all, this dataset represents a critical first benchmark for future development and evaluation of deep learning models for insect detection and identification.

The widespread decline in biodiversity and abundance of pollinator insects is expected to provoke cascading effects on food security and jeopardize ecosystem services crucial for many crops and wild plants. Pollinator monitoring is a crucial element in preventing further decline of pollinators, to which computer vision approaches can make essential contributions. To facilitate research in such approaches, we present a dataset for pollinator detection with accurate annotations. We develop the dataset with an iterative semi-automatic annotation approach, which leverages YOLO to assist with human annotation. We quantify the impact of multiple levels of errors in annotations on training and report the increase in mAP of 28.7% at the final iteration when compared to the manual annotations. Our dataset encompasses pollinator detection for honeybees and bumblebees across various flower treatments over multiple days. Our dataset facilitates the development of deep learning-based methods for automatic large-scale pollinator detection under various real-world field conditions, as well as adjacent computer vision tasks such as small object detection and label correction.

Accurately localizing vocalizing animals from passive acoustic recordings remains challenging in complex outdoor soundscapes. Although automated detection and classification of bird calls have advanced rapidly, reliable spatial localization has lagged, particularly for small arrays of autonomous recorders operating without manual intervention. Here we describe a fully automated pipeline for three-dimensional localization of bird vocalizations using distributed networks of 4 to 6 GPS-synchronized recorders deployed in heterogeneous forest environments. Our framework integrates automated multi-recorder event matching, frequency-selective FFT-based cross-correlation for time-difference-of-arrival (TDOA) estimation, geometric cycle-consistency filtering to resolve ambiguous cross-correlation peaks, and nonlinear optimization of source location and effective sound speed. Applied to multi-year datasets from three field sites, the localizations exhibit strong concordance of localizations with independently known landscape features and species-specific habitat associations. These analyses indicate that small, practical arrays can recover ecologically meaningful spatial structure in complex soundscapes without manual curation. This preprint documents the current state of the system and its performance under realistic field conditions.

O_LIAcoustic monitoring is an effective and scalable way to assess the health of important bioindicators like bats in the wild. However, the large amounts of resulting noisy data requires accurate tools for automatically determining the presence of different species of interest. Machine learning-based solutions offer the potential to reliably perform this task, but can require expertise in order to train and deploy. C_LIO_LIWe propose BatDetect2, a novel deep learning-based pipeline for jointly detecting and classifying bat species from acoustic data. Distinct from existing deep learning-based acoustic methods, BatDetect2s outputs are interpretable as they directly indicate at what time and frequency a predicted echolocation call occurs. BatDetect2 also makes use of surrounding temporal information in order to improve its predictions, while still remaining computationally efficient at deployment time. C_LIO_LIWe present experiments on five challenging datasets, from four distinct geographical regions (UK, Mexico, Australia, and Brazil). BatDetect2 results in a mean average precision of 0.88 for a dataset containing 17 bat species from the UK. This is significantly better than the 0.71 obtained by a traditional call parameter extraction baseline method. C_LIO_LIWe show that the same pipeline, without any modifications, can be applied to acoustic data from different regions with different species compositions. The data annotation, model training, and evaluation tools proposed will enable practitioners to easily develop and deploy their own models. BatDetect2 lowers the barrier to entry preventing researchers from availing of effective deep learning bat acoustic classifiers. Open source software is provided at: https://github.com/macaodha/batdetect2 C_LI

Acoustic surveys of bats are currently limited by the detection range of ultrasound microphones. This makes it difficult to survey bats at height, over water, or in other hard to reach locations. Drones, also known as UAVs or UASs, are becoming more affordable and feature rich, resulting in more uptake in conservation, primarily for aerial imagery. In this paper we address current limitations to acoustic bat surveys by developing three autonomous drones for surveying bats; a plane, quadcopter and boat. All three are capable of moving autonomously from one waypoint to another while carrying a bat detector to record bats and are all low cost (under US$950) with large operational ranges that make them suited to aerial transect work.\n\nInitial testing highlighted ultrasound noise generated by the drones as a major issue for recording bats from these systems. This was mitigated through iterative design of the microphone placement and the vehicles themselves. Subsequent testing in real world settings, in the presence of bats, demonstrated that bats could be recorded under autonomous navigation and that the ultrasound interference could be reduced to a negligible level.\n\nAutonomous drones offer an exciting new tool in the tool box of bat workers and researchers. They allow us to study areas previously inaccessible on foot, as well as at heights that have previously been inaccessible. We also discuss current limitations to this technology, including legal considerations, hidden costs and potential impacts on bat behaviour.

Large-scale biodiversity monitoring is often inhibited by taxonomic obstacles. While deep learning has demonstrated efficacy in species identification, the increasing reliance on large Vision Transformers (ViTs) creates computational barriers that restrict usage to cloud-based infrastructure. Recent foundation models, such as BioCLIP and the Insect-1M framework, require parameter counts exceeding 100M, rendering them unsuitable for edge deployment in field operations. This study presents Elytra 1.0, a computer vision model optimized for edge deployment and capable of classifying 3,127 common North American insect species. The dataset, comprising 2.6 million images, includes all insect species in North America with over 1,000 research-grade observations on iNaturalist. An EfficientNet-B0 architecture was trained using transfer learning from ImageNet with adaptive learning rate scheduling. The model achieved 91.27% Top-1 Accuracy and 97.6% Top-5 Accuracy on an internal test set (N=289,151 images). To rigorously evaluate generalization beyond photographer-specific patterns, an independent observer-excluded test set (N=5,780 images, 578 species) was constructed comprising images exclusively from photographers who contributed zero training data. A post-hoc spatiotemporal audit revealed this test set was heavily skewed toward the Neotropics (Mean Lat: 6.05{degrees} N) during the boreal winter (Dec 2025-early Feb 2026). Despite this significant biogeographic and phenological shift from the predominantly temperate training data, the model achieved 86.68% Top-1 Accuracy (95% CI: 85.8-87.5%). This confirms that Elytra 1.0 relies on robust morphological features rather than learning background environmental correlations, maintaining high performance even in novel ecological contexts.The resulting model file size is 30 MB with an inference speed exceeding 700 frames per second (FPS) on mobile hardware. These results indicate that optimized convolutional architectures can achieve competitive accuracy with server-grade transformers while remaining suitable for decentralized, offline monitoring applications.

Measuring forest biodiversity using terrestrial surveys is expensive and can only capture common species abundance in large heterogeneous landscapes. In contrast, combining airborne imagery with computer vision can generate individual tree data at the scales of hundreds of thousands of trees. To train computer vision models, ground-based species labels are combined with airborne reflectance data. Due to the difficulty of finding rare species in a large landscape, the majority of classification models only include the most abundant species, leading to biased predictions at broad scales. Extending classification models to include rare species requires targeted data collection and algorithmic improvements to overcome large data imbalances between dominant and rare taxa. In addition, large landscapes often require multiple acquisition events, leading to significant within-species variation in reflectance spectra. Using a multi-temporal hierarchical model, we demonstrate the ability to include species predicted at less than 1% frequency in landscape without losing performance on the dominant species. The final model has over 75% accuracy for 14 species with improved rare species classification compared to a baseline deep learning model. After filtering out dead trees, we generate landscape species maps of individual crowns for over 670,000 individual trees at the Ordway Swisher Biological Station within the National Ecological Observatory Network. We estimate the relative abundance of the species within the landscape and provide three measures of uncertainty to generate a range of counts for each species. These maps provide the first estimates of canopy tree diversity within NEON sites to include rare species and provide a blueprint for capturing tree diversity using airborne computer vision at broad scales.

Tropical forest canopies contain many tree and liana species, and foliar and reproductive characteristics useful for taxonomic identification are often difficult to see from the forest floor. As such, taxonomic identification often becomes a bottleneck in tropical forest inventories. Here we present a drone-based workflow to automatically acquire large volumes of close-up, ultra-high resolution photos of selected tree crowns (or specific locations over the canopy) to support tropical botanical and ecological studies (https://youtu.be/80goMEifpc4). Our workflow is built around the small, easy-to-use DJI Mavic 3 Enterprise (M3E) drone, which is equipped with a wide-angle and a telephoto camera. On day one, the pilot maps a forest area of up to [~]200 ha with the wide-angle camera to generate a high-resolution digital surface model (DSM) and orthomosaic using structure-from-motion (SfM) photogrammetry. On subsequent days, the pilot acquires close-up photos with the telephoto camera from up to 300 selected canopy trees per day. These close-up photos are acquired from 6 m above the canopy and contain a high level of visual detail that allows botanists to reliably identify many tree and liana species. The photos are geolocated with survey-grade accuracy using RTK GNSS, thus facilitating spatial co-registration with other data sources, including the photogrammetry products. The primary operational challenge of our workflow is the need to maintain RTK corrections with the drone to ensure that close-up photos are acquired exactly at the predefined locations. The maximum operational range we achieved was 3 km, which would allow the pilot to reach any tree within a [~]2800 ha area from the take-off point. Although our workflow was developed to support taxonomic identification of tropical trees and lianas, it could be extended to any other forest or vegetation type to support botanical, phenological, and ecological studies. We provide harpia, an open-source Python library to program these automatic close-up photo missions with the M3E drone (https://github.com/traitlab/harpia). Data/code for peer review statementWe provide harpia, an open-source Python library to program these automatic close-up photo missions (https://github.com/traitlab/harpia). Drone imagery and labelled close-up photo data are not yet publicly available because they were acquired with the goal of publishing benchmark machine learning datasets and models for tree and liana species classification and prior publication of the data would jeopardize this future publication.

Passive acoustic monitoring (PAM) - the use of autonomous recording units to record ambient sound - offers the potential to dramatically increase the scale and robustness of species monitoring in rainforest ecosystems. PAM generates large volumes of data that require automated methods of target species detection. Species-specific recognisers, which often use supervised machine learning, can achieve this goal. However, they require a large training dataset of both target and non-target signals, which is time-consuming and challenging to create. Unfortunately, very little information about creating training datasets for supervised machine learning recognisers is available, especially for tropical ecosystems. Here we show an iterative approach to creating a training dataset that improved recogniser precision from 0.12 to 0.55. By sampling background noise using an initial small recogniser, we can address one of the significant challenges of training dataset creation in acoustically diverse environments. Our work demonstrates that recognisers will likely fail in real-world settings unless the training dataset size is large enough and sufficiently representative of the ambient soundscape. We outline a simple workflow that can provide users with an accessible way to create a species-specific PAM recogniser that addresses these issues for tropical rainforest environments. Our work provides important lessons for PAM practitioners wanting to develop species-specific recognisers for acoustically diverse ecosystems.