Ecological Informatics

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.

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.

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.

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.

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.

Climate affects biological processes and organisms. It is extensively recorded, and its effects are frequently studied. Climatic variables are recorded as time series of measurements, but these are often summarized as long-term averages, throwing away biologically relevant information. Subsequent analyses can therefore miss important climate effects or suffer from low predictive power. Here, we propose a new method, StatClim, inferring the most informative climatic summary from the data. Instead of taking a fixed arbitrary summary of a time series, StatClim automatically estimates the beginning and width of a climatic window and simultaneously calculates data aggregation, which best predicts a response variable. We implemented the proposed approach as a Bayesian hierarchical model in Stan. We illustrated the potential of StatClim on published data on mast seeding of Fagus sylvatica. StatClim correctly identified known climatic windows relevant for this tree species, but also uncovered that mast seeding is triggered by high temperatures within these windows. We demonstrated flexibility of the model to account for complex data structures, we ran power analysis showing the ability of the model to identify climatic effects even with modest datasets, and we evaluated its false positive rate, which is lower than 0.01. We discuss the advantages and limitations of StatClim. Compared to existing alternatives, it is suitable for hypothesis testing and offers important advances, such as estimation of the type of aggregation. In combination with existing methods, StatClim will facilitate data-driven use of climatic and microclimatic data and therefore discovering so far unknown climatic effects on biological processes and ecosystem functioning.

Global biodiversity loss is a consequence of habitat degradation and fragmentation caused by the expansion of the Anthropocene and the overexploitation of natural resources. Passive acoustic monitoring enables longitudinal assessment of biodiversity but generates massive datasets. Advancements in machine learning have enabled the development of convolutional neural networks capable of automating this process. We developed a lightweight 2D convolutional neural network capable of detecting echolocation calls of the greater sac-winged bat, Saccopteryx bilineata, an insectivorous aerial edge-foraging bat, across a tropical restoration gradient in Para, Brazil. The 2D-CNN model achieved a 97.84% accuracy and 100% precision. Deploying this model across 157,335 field recordings across 29 sampling points resampled over 3 years (2023-2025) yielded 2,098 positive detections of S. bilineata. Generalised linear mixed models revealed that time since reforestation significantly increased the odds of detecting S. bilineata ({beta} = 0.22, p < 0.001), and that detection probability was 2.14 times higher in restoration plots than in forest habitats ({beta} = 0.76, p = 0.0701). Acoustic hardware had a significant covariate influence in both fitted models ({beta} = 0.52, {beta} = 0.72, both p < 0.001). These findings demonstrate the practical applications of deep learning and bioacoustics to quantify habitat recovery and the value of restoration in regenerating functional ecological niches for aerial insectivores during early successional stages.

Habitat destruction leads to a severe decline in biodiversity. Ascertaining effects of habitat loss afterwards is regularly done, whereas predicting the effects of different habitat loss scenarios on biodiversity beforehand is much preferred but as of yet rarely possible. Such predictions would help inform management decisions to prevent biodiversity loss in the first place, or at least mitigate effects if necessary, by providing handles to set priorities in making trade-offs. To this end, we developed and applied an adapted version of the information-theoretical framework of Maximum Entropy Theory in Ecology (METE) to predict species-area relationships in tropical forest landscapes undergoing habitat loss and fragmentation. Species-area relationships provide information about biodiversity and can, in undisturbed systems, be predicted using METE, based on state variables such as the total number of individuals, the total number of species, and the total area size. In disturbed systems, however, it is established that METEs predictions are off. Using an individual-based model, we explore how METEs state variables should be adjusted, based on the landscape characteristics habitat cover and degree of habitat fragmentation, to restore its accuracy of species-area relationship predictions in disturbed environments. We validate our adjusted METE predictions with empirical data. Our adjusted METE predictions demonstrate the severity of fragmentations impact on biodiversity, relative to what could be expected in a pristine, continuous environment. With this new method, we can estimate the immediate effects of habitat loss on biodiversity, as well as its long-term effects.

Plant phenology influences biotic interactions and the cycling of water, carbon, and nutrients, ultimately affecting global climate. Accurately quantifying phenological dynamics is therefore essential for ecological forecasting, yet estimates vary widely depending on data sources, smoothing approaches, and date-extraction techniques. For this reason, the selection of an appropriate processing pipeline requires benchmarking under controlled conditions. In this study, we use a Land Surface Model (LSM) to simulate ecosystem dynamics at four temperate forest sites under five lesser studied climate perturbations. The resulting synthetic dataset allows us to control for environmental conditions while comparing commonly used phenological detection methods (relative thresholds at 10%, 20% and 50%, highest rate of change, detection of inflection points in the curvature angle, detection of recovery and senescence points). Amongst tested methods, we find general agreement in responses to climate alterations, where inflection and low thresholds (10%) methods provide the closest match to the model predicted start of season, with a maximum discrepancy of 4 days. Higher thresholds (20%, 50%) are closer to the model predicted end of season, though discrepancies reach up to 10 days. High thresholds and rate-of-change methods also reliably detect the peak of season. We further show that some methods detect phenological responses missed by temperature-only type growing season, supporting the need for more realistic phenological representations in LSMs. Beyond evaluating extraction methods within a common framework, our study expands on the potential of using LSMs as sandboxes for ecological testing rather than solely predictive tools.

Trail cameras have become ubiquitous tools for ecological data collection over recent decades. Despite progress in the development of automated algorithms and artificial intelligence for image classification, our ability to process large volumes of data remain limited by the need for trained human observers to make refined judgements. We provide guidance on placement of trail cameras for observing Asian elephants (Elephas maximus) and outline a protocol for training and testing naive human observers in performing image classifications (age/sex class and group composition) that cannot yet be automated. This process can be used to develop a high-throughput workflow capable of extracting useful data from large volumes of images. Our training material consisted of 14,007 images collected from 6 trail cameras around Udawalawe National Park in Sri Lanka from 2017-2019. In the first stage, expert observers (n=3) trained a group of inexperienced participants (n=4), who engaged in an iterative process to develop a protocol document. The document was then tested on a second set of subjects (n=6) each of whom classified 350 test images in four separate sequential batches using quantitative measures of precision and accuracy. The test set was sampled from 54,435 images from an additional 25 cameras. When compared to expert observers, they achieved a fair level of precision (Fleiss' kappa = 0.247) and 82.6% accuracy. Our approach can usefully be extended to other species and contexts.

The Siamese fighting fish (Betta splendens) is well-known for its high levels of aggression and complex, stereotyped displays, making it an ideal model for testing deep learning-based tools for high-throughput, unbiased quantification of aggressive behaviour. A deep-learning object recognition model (BettaAI) was developed to detect the species' aggressive displays, including threat (opercula opening and fin distension), attack (bites) behaviours, as well as surface air breathing, a behaviour correlated with the frequency and intensity of aggression. Videos of males confronting their mirror image were acquired simultaneously with synchronised top and side cameras. Two separate models were trained, with some behaviours identified from top-view frames and others from side-view frames; outputs from both models were combined and post-processed for final quantification. BettaAI performed with high accuracy: F1 scores were 0.99 for the top-view model and 0.93 for the side-view model. Average Precision (AP) at 0.5 IoU exceeded 0.98 for all categories except bites, a rare behaviour in the dataset, which reached 0.865. Comparisons with a manually curated "ground-truth" dataset from videos of 10 males showed strong concordance between automated quantification and manual annotations by expert observers across all behaviours. A case study on 461 males reacting either to a mirror or to a video playback of a conspecific demonstrated the high-throughput and unbiased advantages of BettaAI, with threat displays being more frequent toward the mirror image and attack rates not differing between treatments. This work underscores the potential of BettaAI for automated, bias-free quantification of aggressive displays in B. splendens.

Climate change is reshaping weed population dynamics, creating non-stationary management challenges for which static decision rules are insufficient. This chapter presents and evaluates an open surveillance toolkit of artificial intelligence tools for recognizing and mapping Ambrosia artemisiifolia (common ragweed) at complementary spatial scales, tested across two phenological phases (seedling, September 2024; adult, December 2024) in a lentil paddock in central Chile. At the field scale, a YOLOv11 detection model coupled with Slicing Aided Hyper Inference achieved mAP50 of 0.886. Cross-domain experiments revealed that single-site models collapse catastrophically when deployed in new geographies (mAP50 dropping to 0.108), but multi-domain training recovers performance to 0.874. At the satellite scale, PRESTO foundation model embeddings correlated strongly with ground-truth weed density (r = 0.739); geographically weighted regression explained up to 91% of local density variance. All methods are released as the ragweed-ai-toolkit (https://github.com/agroia-lab/ragweed-ai-toolkit), a modular open-source Python package enabling adaptation to new species and geographies.

1. Soil invertebrate biomass is closely linked to ecosystem processes, yet its routine quantification remains constrained by the effort required to weigh large numbers of small individuals. Length-mass relationships provide an indirect alternative, but body length can be difficult to standardise and extract automatically from images of curved, coiled or contracted organisms. Projected body area may provide a more tractable image-derived predictor, but its performance across diverse soil-faunal taxa and at the community scale remains insufficiently evaluated. 2. Here, we developed a framework using projected body area and contour length to estimate fresh and dry mass from measurements of 1,792 soil invertebrates representing 15 higher-level taxa. We fitted taxon-specific models for six well-represented groups and general models across taxa, and evaluated predictive performance using leave-one-out cross-validation. We then assessed cross-dataset transferability by comparing our data with those of Lemoine et al. (2026) and pooled the two datasets to derive combined models. Finally, we tested whether individual predictions could be scaled to independently measured community biomass. 3. Taxon-specific models generally reduced prediction error relative to general models, with mean absolute percentage error (MAPE) values of 6.9-37.3% for fresh mass and 7.6-30.2% for dry mass. General models retained MAPEs of 27.2% and 22.4%, respectively. Comparisons of our dataset with that of Lemoine et al. (2026) showed no detectable differences in six of eight taxon-by-mass slopes. Although intercepts frequently differed, no overall dataset effect was detected after accounting for taxa. Pooling both datasets broadened model development, while taxon-specific models generally retained lower error than the combined general model. Finally, community-level predictions tracked measured biomass across all model sets (R2 = 0.87-0.99), but absolute error varied among models and faunal groups (MAPE = 9.6-52.5%). 4. Body area and contour length extracted from images therefore provide a practical basis for scaling soil invertebrate biomass estimation from individuals to communities. Integration with automated imaging and segmentation tools such as flatbug, BIODISCOVER and CollembolAI could extend soil invertebrate monitoring at larger scales beyond abundance and taxonomic composition to biomass, while providing scalable inputs for analyses of food webs and general ecosystem ecology.

1. Species monitoring programmes regularly encounter missing data, which complicates tasks such as estimating population size or detecting temporal trends. Selecting an imputation method suited to the properties of the data is therefore an important practical challenge, particularly for species exhibiting overdispersion and zero-inflation. 2. We compare seventeen imputation methods, comprising thirteen Poisson-based statistical models (accounting for overdispersion and zero-inflation, with fixed, random and multivariate structures) and four contrast-based approaches (LORI, MICE, missForest, correspondence analysis). Using four complete monitoring datasets of waterbird species surveyed across France and Italy over 21 years, illustrating a variety of abundance distributions, we introduced missing observations under a realistic Missing At Random mechanism, at rates from 5% to 70%. We evaluated all methods on computational burden, prediction accuracy for positive and null counts, quantification of uncertainty, and accuracy of population-size estimates. 3. No single method dominates across all criteria. Statistical models and contrast-based approaches yield similar point predictions, but only statistical models provide a genuine measure of uncertainty. Models that jointly account for overdispersion and zero-inflation perform best at predicting zero counts and achieving reliable prediction intervals, and the most suitable method ultimately depends on the abundance distribution of the target species. 4. These results provide practical guidance for ecologists selecting an imputation strategy for incomplete count data, highlighting the trade-offs between predictive accuracy, computational cost, and the ability to propagate uncertainty through subsequent ecological analyses such as population-trend estimation.

1. Detecting causal relationships from ecological time series remains challenging because community dynamics are often nonlinear, high-dimensional, noisy, and strongly synchronized. These characteristics can reduce the reliability of existing causal inference methods and complicate the identification of ecological interactions in complex natural systems. 2. I introduce Forest-weighted Causal Correlation (FORCC), a machine-learning-based framework that infers directional causality from adaptive neighborhood structures learned by random forests. Rather than evaluating causal direction through cross-map reconstruction, FORCC quantifies directional dependence from the preservation of neighborhood structure using rank-based associations among forest-weight matrices. 3. Simulations of coupled nonlinear systems showed that FORCC accurately recovered causal direction while substantially reducing synchronization-induced false-positive detections relative to conventional convergent cross mapping. Chatterjee's {xi} exhibited the strongest resistance to synchronization-induced errors among the evaluated rank-based statistics. 4. Application to a 12-year marine community time series from Maizuru Bay, Japan, identified asymmetric relationships between individual taxa and community dynamics, revealing candidate driver-like and passenger-like roles that were not readily explained by seasonal synchrony alone. 5. By combining adaptive neighborhood learning with rank-based causal inference, FORCC provides a scalable framework for causal discovery in ecological time series. The method offers a practical tool for screening causal relationships, generating ecological hypotheses, and investigating complex interactions in multivariate natural systems.

Data curation-the process of collecting, cleaning, and joining raw datasets for unified analysis-is a time-consuming yet crucial part of any data science project. With the emergence of powerful agentic coding tools, it is tempting to "vibe curate"-i.e., to prompt AI agents to perform data curation operations and blindly trust their outputs in order to finish work more quickly. However, this can exacerbate two key challenges that already plague manual data curation workflows: reproducibility-the ability to trace the exact set of modifications applied to raw datasets-and verifiability-the ability to audit changes and confirm that data is processed properly. To address these issues, we introduce Scrub Data, a framework that enables the use of AI agents in data curation workflows while transparently maintaining both reproducibility and verifiability. The framework involves an interactive data curation loop between a user and an AI agent, backed by a data provenance graph that tracks and versions all dataset updates. The data graph requires that all updates are formatted as self-contained executable steps, allowing any version of a dataset to be reconstructed from raw data by playing forward the transformations stored in the graph. This workflow takes place within a lightweight web application that is designed to be modified by users and agents in real-time to create custom visualization tools for identifying data curation needs and verifying outcomes. We demonstrate the usefulness of the framework with a real data curation use case from animal movement ecology. Starting with 89M raw GPS coordinates, we use the framework to curate a benchmark dataset of 2.5M coordinates to be used for machine learning or statistical analysis. The framework is available for use and extension at https://github.com/justinkay/scrubdata.

Reliable demographic data are critical for seal population assessment and management, yet remain challenging to obtain from drone surveys without automated tools. Using data from harbor seal (Phoca vitulina) haulout surveys conducted in British Columbia, Canada, in 2023-2024, we developed and tested an automated pipeline that integrates detection, morphometric extraction, and size-class classification of seals into a single system. The pipeline combines cascaded DeepLabV3+ segmentation (water removal, colony-level seal detection, and individual body-contour refinement), principal component analysis (PCA)-based extraction of body length and maximum width, and three-class size-class determination based on body length (pups, juveniles, subadults/adults). We evaluated the pipeline on 2,483 manually annotated seals across four drone platforms (DJI Mavic 3 Thermal, Mavic 2 Pro, Mini 3 Pro and Mini 2), five substrate types, and two sensor modalities (RGB, thermal infrared). Aggregate detection F1 was 0.914. Orthomosaic geometric accuracy, quantified as the RMSE of calibration-rod length on the orthomosaic, ranged from 1.0 cm to 18.0 cm depending on the platform-altitude-substrate combination, a spread that aggregate metrics would mask. Relative to manual annotation on the same orthomosaics, automated RGB body lengths showed near-zero bias (-0.6 cm; RMSE = 12.5 cm), whereas automated thermal lengths were systematically shorter (bias -19.9 cm, RMSE = 24.1 cm), establishing the unsuitability of thermal imagery for automated quantitative pinniped morphometry. For RGB imagery, threshold classification by body length reached an overall accuracy (OA) of 78.7%; for individuals outside the uncertainty zones (defined as {+/-}2 x the geometric RMSE on calibration rods around each class boundary), OA reached 90.9%. Within the RGB subset, OA was 78.9% at 20 m altitude, declining to 75.0% at 40 m where geometric error dominated. The pipeline enables automated demographic monitoring of pinniped haulouts with quantitatively characterized uncertainty across survey conditions.

(1) Kelp forests support ecologically, commercially, and culturally important species, yet their benthic communities remain difficult to monitor consistently across space and time. Scientific SCUBA surveys provide valuable observations but can be constrained by safety, bottom time, depth, and the spatial extent that divers can efficiently cover, motivating complementary approaches for collecting standardized benthic imagery. (2) We developed and field-tested a standardized, open-source framework for benthic visual-imaging surveys using a small, modular remotely operated vehicle (ROV). Across four cost-scalable hardware configurations, we integrated navigation sensors, custom BlueOS Extensions, and our surftrak contribution to ArduSub, which enables fixed-altitude flight above the seafloor, and developed two open-source desktop applications that streamline field operations (rov_flight_ops) and imagery processing (rov_imagery_processing). We also evaluated power supplies, benthic lighting, camera settings, field protocols, and telemetry-processing workflows over 303 transects conducted on 64 survey days from 2022 to 2026. (3) The ROV operated across soft-sediment and rocky habitats, currents of 0-3 kt, and underwater visibility of 0-10 m. surftrak held the ROV within 0.2 m of its target altitude for 96% of 21 h of survey transects, producing an approximately 1 m^2 standardized image footprint. Sufficient illumination and optimized camera settings yielded high-resolution imagery in visibility below 2 m, while topside power increased endurance and operational flexibility. Combining Doppler velocity log and underwater acoustic positioning data through the vehicle's extended Kalman filter improved navigation and post hoc transect reconstruction. During 2026 surveys, rov_flight_ops and rov_imagery_processing automated flight logging, log retrieval, fault diagnosis, and the sorting, development, and telemetry annotation of survey imagery. Dense kelp canopies nevertheless constrained vehicle operation, and downward-facing imagery provided limited observations of cryptic or obscured organisms. (4) We provide configuration-specific bills of materials, hardware and wiring documentation, software and firmware inventories, camera settings, open-source field and imagery-processing software, step-by-step field protocols, and pre- and post-dive checklists to support replication and adaptation of the framework. Together, these resources provide a field-tested starting point for expanding standardized benthic imaging while complementing diver surveys across kelp forests and other shallow coastal ecosystems.