Ecological Informatics

Remote sensing mapping of plant functional diversity (PFD) is increasingly used at the science-policy interface. Yet assessments of uncertainty and quality standards for these products are seldom provided, as the impact of measurement uncertainty on PFD estimates, i.e., the variability across pixels, is still poorly understood. We use trait maps and hyperspectral imagery simulations from the Biodiversity Observing System Simulation Experiment (BOSSE) to assess the impact of different types of uncertainty on the estimation PFD. Results show that uncertainties stemming from random errors affect the estimation of PFD (up to three times) more than those arising from systematic errors, with spectral and spatial error correlation reducing their impact. Standardization computing PFD metrics reduces systematic errors, making in-situ-measured plant traits and remote sensing reflectance factors more robust estimators of PFD than other remote sensing proxies derived from reflectance (e.g., spectral indices or optical traits). The impact of uncertainty propagation from reflectance factors to these often-preferred secondary proxies increases with their complexity. Nonetheless, trait uncertainty can spuriously diminish differences between PFD derived from the ground (plant traits) and satellite imagery, as it can reduce uncertainty-induced or inherent biases between the PFD metrics computed from each trait (plant or spectral). This study provides new understanding of how uncertainty propagates into PFD estimates; this knowledge is relevant for optimizing the selection and defining quality requirements for PFD estimation. It also suggests that random and systematic uncertainties should be quantified separately at the reflectance factor level to guide their use and interpretation in PFD analysis.

Species distribution models are widely used for the evaluation and analysis of species spatial distributions and responses to environmental changes. Due to this, there's a wide diversity of tools and algorithms to perform these types of analyses. However, despite being a mature field of study, SDM still requires a large amount of research intervention and decision-making to produce reliable models. This is a natural consequence of each species unique response to environmental factors, but in practice it hinders the potential of SDMs to be used effectively and with confidence across large sets of species. Here we present the AutoMaxEnt repository, a collection of functions and tools for the automation of maximum entropy species distribution models. AutoMaxEnt offers a flexible environment to model species distribution models, automating most of the data pre-processing and preparation and giving the user the possibility of exploring multiple model scenarios automatically or with control over how those scenarios are generated. AutoMaxent offers model fit control, multiple background points generation algorithms, automatic model and variable selection, and study area configuration, among other parameters. AutoMaxEnt allows for the automation of model fitting, evaluation and model selection across different species, study areas and time periods.

Risk identification of non-native species is widely used to support prioritization, early warning, horizon scanning and management decisions. Over the past two decades, the Invasiveness Screening Kit (ISK) framework of decision-support tools has developed through aquatic, terrestrial animal and terrestrial plant applications, but until now these toolkits have been distributed and used as separate software environments. ISK v3 provides a single integrated Microsoft Excel/Visual Basic for Applications platform for the Aquatic Species Invasiveness Screening Kit (AS-ISK), Terrestrial Animal Species Invasiveness Screening Kit (TAS-ISK) and Terrestrial Plant Species Invasiveness Screening Kit (TPS-ISK). It preserves the established questionnaires, scoring logic and multilingual implementation of the three toolkits within a single integrated interface, while introducing improved harmonized database workflows, screening record management, taxonomic verification, threshold handling with built-in calibration, risk summaries, reporting, merging and export functions as well as controlled access to separate Excel instances. The platform supports 31 languages, toolkit-specific databases, conversion of compatible v2 databases, import of legacy aquatic first-generation ISK databases into AS-ISK, online verification through major taxonomic and biodiversity data resources, and use of a global a priori categorization dataset for calibration. ISK v3 preserves the reproducible screening structure of the three toolkits while improving consistency, transparency and traceability in non-native species risk identification and reducing fragmentation across toolkits. It provides a common platform for researchers, managers and institutions applying risk identification workflows across aquatic, terrestrial animal and terrestrial plant taxa.

Coral reef fish assemblages are widely used as indicators of ecosystem condition, yet manual annotation of underwater video remains a major bottleneck for scalable biodiversity monitoring. Despite rapid progress in automated detection, ecologically realistic and publicly available datasets remain scarce, particularly for the Western Indian Ocean. Here, we present WIO-ReefFish, a reef fish detection dataset derived from diver-operated line-intercept transects and designed for ecological monitoring under natural survey conditions. WIO-ReefFish comprises 1,000 ultra-high-definition images (3840 $\times$ 2160 pixels) and 6,768 exhaustive bounding-box annotations spanning 24 taxonomic categories, thereby preserving full-frame assemblage structure in complex reef scenes. We also establish a standardized benchmark across nine object detection models under two complementary protocols: class-aware detection and class-agnostic fish localization. Detection performance was consistently higher under the class-agnostic protocol. The best-performing model (RT-DETR) improved from 0.48 mAP50 in the class-aware setting to 0.70 mAP50 when taxonomic constraints were removed, indicating that taxonomic discrimination remains substantially more challenging than fish localisation in reef imagery. Spatially independent evaluation revealed a pronounced generalisation gap, particularly for taxonomic detection, whereas class-agnostic fish localisation remained substantially more robust across transects and countries. Together, these results establish WIO-ReefFish as a realistic benchmark for automated reef fish detection and provide a foundation for more robust computer-vision tools in coral reef biodiversity monitoring. The WIO-ReefFish dataset and associated benchmarking resources are publicly available.

Insects are a key group in ecosystems and economics, but are increasingly under threat, making ongoing monitoring essential to preserving them. Traditional monitoring approaches are time consuming, among other limitations. Automated acoustic recognition can be a good solution for monitoring arboreal, soniferous insects, such as cicadas, and the acoustic biology of cicadas in Australia is understudied, especially using automated methods. To evaluate the applicability of automated approaches for detecting cicadas, we used the deep-learning acoustic model BirdNET (v2.4) to perform embedding searches for Cystosoma schmeltzi (lesser bladder), Illyria burkei (eastern rattler), Macrotristria intersecta (corroboree cicada), Thopha sessiliba (northern double-drummer), and Pauropsalta opaca (fairy dust squawker), using audio recorded by Passive Acoustic Monitoring over 15-months at 18 sites in north Queensland. All classifiers achieved high area under the precision-recall curve (AUPRC) ranging from 0.753 to 0.982. We used a precision-prioritised threshold ([≥] 0.85) to constrain the false positive rate while retaining adequate outputs, which were consistent with the documented species ecology and life history. This approach was readily applicable to the Cicadidae, offering a scalable solution for biodiversity monitoring in complex acoustic environments.

Early detection of invasive species is critical to efficiently managing biological invasions. Passive acoustic monitoring combined with deep learning has become an effective tool for identifying invasive anurans from their species-specific mating calls. In practice, however, these systems still depend on manually retrieving recordings and running inference on a dedicated workstation, which delays detection at invasion fronts, especially on remote islands where frequent travel is impractical. Here, we present a lightweight convolutional neural network for detecting Rhinella marina and Polypedates leucomystax that can be used with low-power microcontrollers, enabling continuous, real-time monitoring of these invasive species. Our model achieved a mean invasive-species F1 of 0.87 with precision higher than 0.88 for both species, comparable to the performance of substantially larger models at a fraction of the model size. Cross-domain evaluation confirmed reliable detection on two out-of-training locations: Iriomote Island, Japan, and on an independent dataset from Australia, demonstrating generalization across recording conditions and biogeographic contexts, albeit with local threshold calibration for P. leucomystax at Iriomote. Our results demonstrate that effective acoustic detection of invasive species is feasible on self-contained, battery-powered devices, removing the dependence on server infrastructure and enabling autonomous early warning systems in remote protected areas.

1. Passive acoustic monitoring (PAM) is increasingly used for ecological research, biodiversity monitoring, assessment, and reporting. Automated species classifiers make it feasible to process large audio datasets but generate numerous detections that often need validation before use in downstream analyses or formal outputs. 2. Method development in PAM has focused on classifier building and downstream models that account for imperfect detection, yet the practical step between these - post-classification validation - remains weakly supported and is often implemented through ad hoc workflows. This increases manual handling, creates scope for transcription or consolidation errors, limits transparency and makes it difficult to document what was reviewed. 3. We introduce PAMalytics, an open-source, no-code, local browser-based application to support post-classification validation as a standardised workflow stage. PAMalytics ingests detections from any classifier, allows users to define how detections are sampled for review, and presents selected detections alongside their spectrograms with audio playback in one unified interface. Sampling strategy and review decisions are tracked alongside reviewer identity improving traceability and reproducibility across the validation workflow. 4. Case studies with Conservation International Cambodia and Imperial College London demonstrate PAMalytics in two validation settings. In Cambodia, gibbon predictions from a large, uneven dataset were sampled within sites, with likely classifier errors prioritised for validation. At Imperial, Amazon bird detections were sampled across each species classifier-confidence range before biodiversity metrics were derived. In both cases, PAMalytics reduced manual handling and validation time. By turning an ad hoc step into an accessible, structured workflow for conservation practitioners, PAMalytics fills a practical gap in the PAM bioacoustics pipeline and strengthens the link between automated detections and evidence used in biodiversity monitoring and reporting.

Global Biodiversity Information Facility (GBIF) occurrence retrievals for an irregularly shaped region are limited by the API spatial query capabilities - rectangular envelopes or size/vertex-limited WKT polygons - neither of which conform to protected areas, sacred groves, wetlands, panchayat or municipal boundaries or any other arbitrary KML polygon of interest queried by users. This paper presents and validates an open, self-contained, adaptive spatial-tiling protocol that (i) ingests any KML polygon of any shape, size and location on earth, breaks it into a set of GBIF API-compatible rectangular tiles, (ii) queries, cleans and clips the individual records to the target polygon, and (iii) summarises the inventory with a generic diversity-completeness-rarefaction module, with minimal manual re-parameterisation between sites. The protocol implements an iterative quadtree refinement algorithm that adapts tile number, size and location to the target polygon geometry, is combined with a fault-tolerant pagination/retry query system, a boundary-exact two-step clipping procedure and a Chao1-based completeness assessment to ensure statistical comparability between sites of different spatial extent and sampling intensity. The algorithm is implemented in open R source (sf, terra, rgbif, tidyverse) with the tiling algorithm controlled by the four parameters only (initial cell size, area floor, tile overlap threshold, recursion limit), with default settings on a new site by simply changing the input file path. This paper describes in detail its five main components - (i) polygon input and validation, (ii) quadtree adaptive tiling, (iii) polygon coverage verification, (iv) tile-wise GBIF query with retry/shrink pagination and partial data retention, (v) boundary-exact deduplication, clipping and diversity estimation. A downstream generic module estimates diversity, Chao1 richness/completeness and Hurlbert rarefaction, for each taxonomic rank and generates rank-ordered diversity tables as output. The generalisability of algorithm to multiple sites has been demonstrated with second polygon (Sonamukhi Sal forest dominated stretch, Bankura district, West Bengal; approx. 610 sq km) that differs from the first (Bishnupur Sal forest dominated stretch; 938 sq km) in both size and complexity (10 vs 34 KML vertices) and report the tiling and diversity metrics comparable results across the two polygons. With no parameter changes, the algorithm generated 135 adaptive query tiles for Sal forest dominated stretch adjoining Bishnupur, and 86 tiles for Sal forest dominated stretch Sonamukhi SDFP, covering completely the area of both polygons. The number of tiles per 100 sq km is comparable between the two runs (14.4 vs 14.1 tiles) despite the 35% difference in polygon size and 3.4x vertex count. The tile-wise querying with retry/shrink pagination retrieved 6,169 GBIF records (excluding errors) with boundary-exact clipping across 404 species for Bishnupur and 1,222 GBIF records (excluding errors) across 271 species for Sonamukhi; the generic diversity module processed the records without further parameter changes and generated comparable metrics for each rank at both sites. The protocol addresses a general bioinformatic challenge in polygon-based GBIF queries, is provided as an open, reusable, documented method which has been validated on two sites. Because the protocol has so far been validated on only two polygons that differ markedly in size, shape and observer regime, it may be regarded as an initial cross-site validation rather than a comprehensive benchmark, and recommend testing on a broader, globally distributed set of polygons before the approach is treated as a general-purpose standard.

Accurate ecological forecasting is increasingly essential for understanding and managing ecosystem responses to climate variability and anthropogenic pressures; however, prediction remains difficult in lakes because key biological variables, such as phytoplankton biomass, exhibit nonlinear dynamics, observational noise, data sparsity, and non-stationarity. This study aimed to test whether large pre-trained time-series foundation models can improve forecasts of lake phytoplankton dynamics, focusing on chlorophyll-a (Chl.a) and water-quality variables in two ecologically contrasting Japanese lakes, the deep-stratified Lake Biwa and the shallow nutrient-rich Lake Kasumigaura. Using extensive monthly monitoring records spanning up to 30 years, we benchmarked two time-series foundation models - Transformer-based Chronos-T5 and probabilistic Lag-Llama - against 15 statistical (AR, ARIMA, SARIMA, Prophet), machine learning (Random Forest, XGBoost, KNN, SVR), and deep learning (LSTM, CNN, TCN, and SSA-hybrid) approaches. Chronos-T5 achieved superior reproduction accuracy and stability across diverse environmental variables and outperformed all other tested models. This performance stemmed from its ability to capture long-term dependencies and complex temporal patterns through large-scale pre-training. We further identified an optimal training window of approximately 14 years, balancing data sufficiency with regime stability, beyond which model accuracy diminished owing to ecological regime shifts. Our findings highlight the transformative potential of time-series foundation models for ecological forecasting, providing a scalable, data-driven framework that can underpin robust early warning systems and adaptive management strategies for aquatic ecosystems worldwide.

Wildlife re-identification (re-ID) is a widely used and powerful tool with diverse applications in animal ecology and conservation. Current automated methods typically operate on single images of a single body part of the animal. However, a single encounter may contain multiple images capturing different body regions, each providing complementary individual-specific information. In contrast to automated approaches, researchers often manually select the most suitable images and regions for identification based on factors like visibility, occlusion and image quality. This creates a mismatch between automated methods and field practice, limiting the practical adoption of current automated re-ID pipelines. Here, we address this by introducing an encounter-based, multi-body-part re-ID framework, using sea turtles as a model taxon. Our framework combines three elements: (1) An orientation-aware deep learning model, TurtleDetector, that in addition to the full bodies, it also automatically segments key body regions, i.e. heads, front and hind flippers, from images within an encounter; (2) a hybrid body-part-specific retrieval method, that sequentially combines a fast global-feature model (MiewID or DINOv3) with a more accurate but costlier local-feature model (ALIKED with LightGlue); and (3) a merged identity-prediction strategy that selects the highest calibrated similarity score across all available body parts and images of an encounter. We evaluate the framework on three long-term re-ID datasets spanning three species, loggerheads, greens, and hawksbill turtles, under an evaluation protocol that mirrors real-world, time-aware re-ID workflows. Across datasets, combining multiple body regions consistently improved identification performance over the best-performing single body region, resulting to an increase of 4-6% in top-1 accuracy. Interestingly, body regions traditionally underused in sea turtle re-ID, such as the hind flippers and carapaces, provided complementary identifying information that improved encounter-level re-ID when integrated through the hybrid retrieval method. Our findings demonstrate that automated wildlife re-ID can benefit from moving beyond single-image, single-body-part identification towards encounter-level integration of all available visual evidence. Our work further suggests that, where feasible, field photo-acquisition protocols should aim to capture multiple informative views of an individual during each encounter. Importantly, many species and taxa, including elephants, primates, cetaceans, and other large vertebrates, possess such individual-specific features across multiple body regions, highlighting the broad potential applicability of our framework.

Plankton plays a fundamental role in aquatic ecosystems, influencing biogeochemical cycles and serving as a key food source for many organisms. Recent high-throughput imaging technologies enable the rapid acquisition of large volumes of microscopic images, creating new opportunities for monitoring planktonic ecosystems. However, the manual processing and annotation of the vast amounts of data generated by these devices remain time-consuming tasks. In this context, machine learning-based classification models offer a promising solution. In this data paper, we introduce a new labeled freshwater plankton dataset comprising approximately 88,000 images distributed across 43 taxa. We also present the labeling assistance method we used to facilitate dataset annotation. Finally, we present a baseline based on a convolutional neural network (CNN), which achieves a classification accuracy of 93% on our dataset.

1. Re-identifying individual animals (re-ID) from their calls with machine learning (ML) could extend individual-based monitoring to species and areas hard to survey by capture and marking, and published studies often report re-ID accuracy above 90%. Two methodological practices can inflate these values: (1) attributing calls to individuals by territory, nest, tag or station, tying each animal to its own recording conditions, and (2) randomly splitting individual clips into training and test sets, allowing recording context to be shared between them. Each effect has so far been documented only in single-species studies. 2. We implement four diagnostics of how methodological choices affect re-ID accuracy: (1) splitting clips by recording session; (2) background tests using ambient recordings without the animal; (3) within-recording permutation and paired-clip tests where animals share recordings; and (4) open-set tests with thresholds set on other individuals. We applied them to 13 public datasets from nine species (eight birds, one bat), using BirdNET, 19 other neural networks trained on animal or general sound and 16 trained on human speech, each with one classifier. 3. Under session splits, BirdNET re-ID accuracy ranged from 0.16 to 0.97. Within one year, ambient sound alone carried substantial identity information for territorial chiffchaffs (BirdNET balanced accuracy 0.71, against 0.79 for song); background identification exceeded chance for 35 of 38 embeddings and controls. Count-matched random clip splits raised mean accuracy by 0.20 to 0.51 in 18 comparisons across three species. Where several animals shared recordings, only one of three datasets showed clearly that re-ID rested on the animal rather than the recording. With unfamiliar animals present and thresholds calibrated to target 10% stranger acceptance, across 36 neural networks and two scoring methods on wild session-split datasets, the highest median share of known birds' calls accepted and correctly named was 46%, using nearest-clip scoring with four enrolled birds; test-stranger acceptance could exceed the target. No neural network beat BirdNET on every dataset, and network rankings depended on layer choice. 4. The diagnostics need no new models and apply wherever a dataset holds what they require: session labels, background recordings, shared recordings or enough individuals. Re-ID accuracy reported without them can substantially overstate what acoustic re-ID achieves, and should not be used to plan population monitoring until they are applied.

Large language models (LLMs) can ease the work of screening titles and abstracts for systematic reviews, but obtaining reliable results requires researchers to make practical choices about which LLMs to use, how to combine their scores into a ranking, and how far down that ranking to read. We aimed to identify a general-purpose workflow that screens accurately, minimises human review effort, and generalises across environmental literature corpora. We ran an ensemble of five open-source LLMs across ten human-annotated systematic reviews from the field of ecology and environmental science spanning 19,777 studies. We then asked: (1) how well an ensemble of LLMs ranks relevant papers above irrelevant ones, and (2) where a human reviewer should stop working down that ranked list. A four-LLM ensemble chosen without any labels came close, on every review, to the best ranking achievable with that reviews annotations (mean Average Precision 0.64 versus 0.66). We tested different rules for when to stop human review, finding the SAFE stopping rule recovered[≥] 95% of relevant records on all ten reviews while requiring a human to screen 55% of the corpus on average. The paper offers a complete workflow that can be adopted for new, unlabelled reviews, using open-source LLMs small enough to run on a high-end consumer laptop, and we provide it as an open-source R package.

Anticipating and managing the impact of herbivorous pests requires evidence-based knowledge about the interactions between pests and their natural enemies. Knowledge of species' ecologies and biological interactions are typically disseminated in unstructured text across hundreds of thousands of scientific articles spanning a rapidly growing scientific literature, but manually screening this literature is time-consuming and impractical at scale. Taxonomic and geographic coverage thus remains sparse, and structured data on interactions between individual species are lacking, particularly for invertebrates. Large language models (LLMs) have recently emerged as a novel and powerful text-mining tool for knowledge extraction and evidence synthesis in a wide range of scientific disciplines, including ecology and conservation. In this study, we use an open-weight, pre-trained LLM to analyse abstracts of a corpus of published material on biological pest control to obtain a global Database of Pest Herbivores and their Natural Enemies (DAPHNE). This method resulted in the extraction of 175,404 species interactions from 112,830 published pest control studies, containing16,755 unique animal taxa resolved at the species level. Interactions include herbivory (granivory, frugivory and gall-formation), predation, parasitism and hyperparasitism, and comprise a range of additional information, such as species' taxonomy, pest status, pest importance, natural enemy importance, provision of biocontrol, associated plants and industries, invasiveness and the vectoring of pathogens and diseases. Validation of the results shows that most data columns were extracted with high recall and precision (>90%) and comparison of DAPHNE with other sources shows a high degree of agreement, but also highlights that the dataset contains information that may have been missed by other sources. In addition to the identification of important crop pests and their natural enemies, DAPHNE may be used for a variety of use-cases, including the identification of invasive species and pathogen vectors, trophic network analysis and literature search and discovery.

The ability to model the complex and uncertain population dynamics of endangered species has improved dramatically in recent decades. However, approaches to identify optimal decisions often require a simplified representation of population dynamics. This leads to a conundrum where managers may be unsure about the output of dynamic decision models because they rely on simplified assumptions of the underlying population dynamics. Here, by pairing integrated population models (IPM) that synthesize diverse ecological data with deep reinforcement learning (DRL) capable of optimizing decisions with high-dimensional uncertainty, we introduce a framework that delivers data-driven and ecologically detailed adaptive management strategies. We demonstrate its utility through application to the supplementation program for the endangered Rio Grande silvery minnow. Using our IPM-DRL framework, we developed an adaptive decision model that selects production and distribution decisions of the supplementation program in response to the observed demographic, hydrological, and genetic environment. The decision model outperformed all heuristic approaches in the simulation across management objectives that weighed persistence and effective population size-related genetic impact differently. For example, the currently deployed supplementation strategy performed 5.3% worse than the decision model under the persistence-focused objective scoring and 185% worse under the genetics-focused one. Analysis of the models decisions in relation to demographic and environmental covariates revealed that minimum sub-population size and total population size were primary drivers of the models decisions. The results demonstrate that the IPM-DRL framework offers a high-performing and interpretable decision-support tool for managing endangered species. SignificanceConservation problems, like imperiled species management, are often challenging because the system dynamics are complex and uncertain. We demonstrate how combining an integrated population model that infers key demographic processes from noisy ecological data with a deep reinforcement learning framework that optimizes management actions addresses these challenges by generating high-performing supplementation strategies for a conservation-dependent species. Our approach embeds two decades of monitoring data within a multi-objective decision-making environment that accounts for ecological uncertainty. The result is a generalizable framework that links ecological inference directly to actionable policy outcomes, enabling scientists and managers to move beyond describing system states and processes toward identifying optimal management actions.

Forest restoration has become a widespread response to deforestation and biodiversity loss, particularly across the tropics. Monitoring restoration is essential for tracking progress and requires accurate knowledge of baseline communities, with threatened species often used as recovery indicators. Passive Acoustic Monitoring (PAM) combined with AI classifiers is enabling this at scale. However, such tools remain immature and may underperform for rare, threatened species underrepresented in training data. Here, we demonstrate the potential of PAM assisted by the AI classifier BirdNET+ V3.0, a January 2026 beta release, to support monitoring of threatened tropical birds. By validating 995 detections from 8,940 hours of audio across a reforestation landscape in Para in the Brazilian Amazon, we found that BirdNET+ confirmed the presence of ~43% of the IUCN Red List threatened species possible in the region, five of which were the first confirmed records for the study site. We also found that this method outperformed traditional survey methods. However, BirdNET+ identified species with mixed reliability, and we could also not determine whether restoration has had any impact on them. Our results demonstrate that PAM with BirdNET+ is a viable, practical method for detecting threatened birds in tropical forest restoration landscapes, but that conventional validation approaches are not appropriate for such species, for which extensive validation effort is an essential cost. Under circumstances where this is not possible, we demonstrate validating the single highest-confidence detection per occasion as a practical alternative.

Protecting deep-sea sessile communities as Marine Protected Areas (MPAs) requires accurate estimates of their spatial distribution and abundance. Existing computer-vision detection methods require careful calibration to avoid counting errors; multi-object tracking (MOT) theoretically offers a more robust alternative. In practice, however, poor lighting and turbidity limit the efficacy of purely kinematic MOT algorithms in deep-sea environments. At the same time, the intra-species visual homogeneity typical of deep-sea scenes means traditional per-frame appearance features are often too weak to maintain stable identities in deep-association MOT algorithms. In this paper we introduce JEPA-SORT, an MOT algorithm that injects V-JEPA 2.1 embeddings into detection-to-track association, to show that video world-model embeddings - which capture not only physical appearance but spatiotemporal trajectory across viewpoints - can help maintain stable identities on visually homogeneous specimens. We provide a benchmark of twelve ROV transects surveying deep-sea sessile communities and evaluate JEPA-SORT against standard MOT algorithms across a range of simulated optical environments. We also propose the Biodensity Utility Metric (BDU), which measures tracker performance in the context of downstream ecological modeling. JEPA-SORT outperforms all benchmarked trackers on IDF1, MOTA, and BDU and is competitive on specimen-count accuracy. Ablations indicate its advantage stems primarily from the discriminative power of world-model embeddings on visually homogeneous footage.

Understanding the consequences of global change for our ecosystems requires reliable, accurate, and scalable tools for detecting ecological change. To this end, recent efforts have focused on scaling up the quantification of demographic processes from individual organisms (the fundamental ecological unit) to the larger spatial scales they inhabit to yield more precise and informative estimations of ecological change, especially in response to disturbance. Here, we describe a transferable workflow that combines robust underwater photogrammetry techniques with machine-learning-assisted image analysis to quantify demographic changes in coral colonies (growth, stasis, partial mortality, and complete mortality) through time on the reef landscape. We then demonstrate several analytical applications of this workflow that can be broadly applied to sessile and/or clonal organisms in other ecosystems. First, we tracked the fates of 3,162 coral colonies through a marine heatwave that caused substantial mortality. Second, our approach enabled detection of size- and taxon-dependent patterns in demographic changes from before to after the heatwave. Finally, for a subset (425) of these colonies, we further demonstrated how our method can be used to track fates across multiple years, including a year of change before and after the heatwave took place, revealing nuanced patterns of survival and mortality, as well as changes in size structure that occurred before and after the disturbance. Our workflow offers an important advance for increasing the accuracy of change detection on coral reefs and other ecosystems by linking demographic changes in individual organisms with spatially extensive image-based surveys through time.

Individual encounter histories are central to capture-recapture models, but fisheries monitoring is often reduced to counts that do not account for variation in detection probability. This study presents a machine-vision workflow for converting long-term video surveillance of Atlantic salmon (Salmo salar) and sea trout (Salmo trutta) spawning runs into individual-level data for capture-recapture analysis. The workflow detects fish-positive frames, saves video clips, extracts fish-head regions of interest, and organizes cropped images for re-identification. A binary EfficientNetB0 fish detector achieved validation accuracy of 0.9918 and PR AUC of 0.9992; at a conservative threshold of 0.98, validation false positives were eliminated while retaining 94.3% recall. A YOLOv8n model localized fish-head regions, and an EfficientNetB0 ArcFace model trained on head images from 700 identities achieved 99.80% accuracy among accepted known matches and an image-weighted false-accept rate of 0.81% when non-training identities were treated as unknown. Zero-shot closed-set retrieval achieved 99.17% top-1 accuracy across 1,416 identities excluded from model training, demonstrating strong generalization to previously unseen identities.

Once central challenge in adopting ecosystem-based fisheries management (EBFM) is understanding and predicting the interconnected processes, feedbacks, and temporal dependencies that characterize ecosystem dynamics. Structural equation models and structural causal models (SEM and SCM) quantify interconnected relationships and evaluate causality from observational data but cannot readily accommodate the incomplete and irregularly sampled data typical of complex systems. Dynamic structural equation models (DSEM) estimate SEMs but accommodate lagged effects, missing data, and non-linear and non-stationary effects. However, there is little guidance regarding how DSEM can be used to support EBFM, particularly within governance and decision-making contexts. Here, we present three case studies that highlight distinct inferential goals for understanding ecosystem processes in the context of EBFM. First, we quantify ecosystem-level indicators to support monitoring of climate-driven ecosystem changes. Second, we examine whether hypothesized direct and indirect ecosystem linkages improve predictions of recruitment. Finally, we compare support for hypotheses of causal drivers of fish productivity. To provide practical advice, we synthesize lessons learned into a workflow to for others navigating DSEM. We suggest that the first step is identifying the inferential objective, which establishes the foundation for the remainder of the workflow, guiding the development of graphical models, model specification, and the selection of statistical and scientific validation steps (e.g., testing conditional independencies) needed to evaluate hypothesized causal relationships. By integrating observational data, expert knowledge, and mechanistic hypotheses within a unified framework, DSEM is a powerful tool for advancing EBFM, and more broadly, understanding the processes that govern complex ecosystems.