ABSTRACT Occupancy models applied to camera trap data are widely used in wildlife monitoring, yet recommendations on the number of survey sites often assume homogeneous conditions and simple model structures. These assumptions are rarely met, and violations can yield biased or imprecise estimates. We used a simulation‐based power analysis to evaluate site number requirements for accurate occupancy estimation under three scenarios of spatial variation in occupancy (𝜓) and detection ( p ): (i) no variation (benchmark), (ii) simple variation (one covariate per process) and (iii) complex variation (multiple covariates, interactions and categorical predictors). Across combinations of baseline occupancy (𝜓 = 0.20, 0.50, 0.80), cumulative detection probability ( p * = 0.65, 0.80, 0.95) and varying degrees of spatial variability, we fitted single‐season occupancy models to simulated datasets of 10–1000 sites. We then identified the minimum number of sites required to achieve root‐mean‐squared error (RMSE) ≤ 0.10 in occupancy estimates. Sample size requirements varied widely across scenarios. Cumulative detection probability strongly influenced required sample sizes, while occupancy mainly mattered when detection was low. With no variation, accurate estimates required 10–40 sites. With simple variation, spatial requirements ranged from 20–120 sites regardless of the degree of spatial variation. Complex variation imposed the highest costs: Models with 2, 3, 4 and 8 parameters per component (𝜓 and p ) required 40–600 sites. Our results highlight that there is no one‐size‐fits‐all number of sites, as requirements differ by an order of magnitude across realistic scenarios. When occupancy and detection vary across space, especially with multiple or categorical predictors, more sites are needed, and sampling effort scales with model complexity. Logistic constraints may force a trade‐off between complex models whose inference is unreliable due to insufficient numbers of sites and simpler, potentially unrealistic models. We recommend simulation‐based power analyses to navigate this trade‐off.
ABSTRACT Altered inundation regimes are reshaping floodplain wetlands and degrading wintering habitats for herbivorous waterbirds, yet long‐term, spatially explicit assessments that link hydrological change to habitat dynamics remain limited. Here, we developed a remote sensing‐based deep learning framework to reconstruct wintering habitat dynamics of herbivorous waterbirds in Dongting Lake from 2000 to 2021. The framework integrates multi‐source environmental predictors, receptive‐field optimization for spatial–context recognition, and Shapley Additive Explanations (SHAP) analysis for ecological interpretation. Results revealed marked habitat decline over the study period, with total habitat area decreasing by nearly half. Habitat loss was highly uneven, with the most severe declines occurring in South (−77.3%) and West (−89.2%) Dongting Lake, while East Dongting Lake remained comparatively stable. Landscape metrics further indicated progressive habitat contraction, increased isolation, and structural simplification, reflecting weakened spatial connectivity and reduced ecological resilience. Among all predictors, water recession timing emerged as the most influential predictor of habitat suitability, with recession in mid‐to‐late October associated with the highest habitat probability. A comparison with the NDVI time series showed that this recession window coincided with higher vegetation productivity. These findings demonstrate that wintering habitat dynamics are strongly associated with hydrological timing, particularly water recession phenology, rather than static water extent alone. The proposed framework provides a transferable approach for identifying hydrological controls on waterbird habitats and offers a scientific basis for adaptive wetland management along migratory flyways.
ABSTRACT The growing availability of ecological imagery creates new opportunities to monitor biodiversity and ecosystem change at broad scales. Yet image classification models remain constrained by uneven sampling, limited labeled data, and differences between training and deployment imagery. Synthetic images generated with generative methods may help address these limitations, but their value across image domains remains poorly tested. We evaluated whether synthetic imagery improves genus‐level classification for 20 North American tree genera using iNaturalist community‐science photographs and Auto Arborist street‐view images. We generated domain‐specific synthetic images with Stable Diffusion v1.5 and trained ResNet‐50 classifiers on controlled mixtures of real and synthetic data. We tested synthetic replacement, supplementation under data scarcity, conventional augmentation alternatives, and transfer between image domains. Replacing real images with synthetic images reduced performance on held‐out real test sets, although declines differed between domains. Adding synthetic images improved classification when real classifier‐training data were scarce, but gains diminished as real data increased. In cross‐domain tests, synthetic supplementation improved some scarce‐data baselines but did not close the transfer gap or replace real‐image training sets. Synthetic images were therefore most useful as supplements rather than direct substitutes. For ecological applications, generated images may support model development for poorly sampled taxa, but representative real images remain essential for robust model generalization.
ABSTRACT Bird monitoring techniques have evolved substantially in recent years. While the most common protocols remain the point count and transect survey, new technologies such as passive acoustic monitoring (PAM) are becoming increasingly popular. As more monitoring programs adopt PAM, it is crucial to understand how comparable it is with classical monitoring derived data. We compared diversity indices (i.e., species richness, Shannon diversity and Jaccard index) derived from both classical (by means of observer‐based point counts) and passive acoustic (by means of recording devices) monitoring and developed a novel estimated Shannon index based on PAM, using species‐specific confidence thresholds and detection rates. We applied the protocols at 126 sites in a heterogeneous landscape within the Italian Alps, enabling us to investigate for the first time whether the effectiveness of different monitoring schemes depends on the habitat type. We show that PAM captured approximately twice as many species as the classical point count method. The species richness estimated by both protocols was strongly correlated (r s = 0.71), but with major differences between habitat types. The difference was greater in diversity hotspots such as wetlands and meadows and lower in alpine habitats. The novel PAM‐based Shannon index was very reliable and yielded similar results to the observer‐based Shannon index, illustrating its effectiveness. Species composition differed significantly between the two protocols (average Jaccard index = 0.38). Although this study highlights the additional information that PAM can provide, bird counts remain essential as they offer further insights (e.g., breeding status) and not all species can be captured by PAM (e.g., acoustically indistinguishable). As point counts are time‐ and cost‐efficient, they will continue to play an important role in bird monitoring. Our results highlight the strengths and weaknesses of each approach and provide recommendations for future monitoring programs.
ABSTRACT Bomb fishing is recognised as the most destructive fishing practice that can be performed in our oceans. This practice involves the use of improvised explosives to stun or kill fish for harvest. Bomb fishing is primarily undertaken on highly threatened tropical coral reef habitats where it accelerates ecological degradation that destabilises communities dependent on these habitats. However, the true extent of bomb fishing in impacted regions is unknown, making it difficult to identify hotspots and activity patterns required to inform management. Here, we present the first quantitative estimate of bomb fishing intensity and temporal patterns in the Spermonde Archipelago, South Sulawesi, Indonesia—a global marine biodiversity hotspot. We developed a machine learning‐based acoustic monitoring system (BombDetect), with an estimated detection distance of at least 16.95 km, and deployed this over a 16‐month period to measure bomb fishing incidents. From the 3698.88 h of audio gathered, we detected and verified 3,567 blasts, equivalent to 23.14 per day, equating to 8446.1 blasts and up to 16.72 ha of reef degradation per year. Bomb fishing occurred year‐round, mostly during daylight hours, with significant reductions on Fridays (local prayer day) and during a period of stormy weather. These findings quantify patterns of occurrence of a dominant driver of long‐term reef decline in the region. We demonstrate the efficacy of a low‐cost approach for quantifying this underreported destructive fishing pressure at scale, which can be readily adapted to other impacted regions to support urgent conservation action.
ABSTRACT Video surveys from drones or planes equipped with cameras have become invaluable tools for monitoring populations of rare and elusive megafauna over increasingly large spatial scales. To analyze hours of such video surveys, automated species detection approaches based on deep learning are now widely used, but simply tallying detections on the successive frames of videos leads to double counting of individuals, significantly biasing population size estimates. Here, we leverage a multiple object tracking approach that links detections across video frames to derive counts of unique individuals. Our tracking‐by‐detection framework integrates the full pipeline from aerial video surveys to automated individual count and estimation of minimum population abundance. Applied to aerial video surveys of dugongs ( Dugong dugon ) on the west coast of New Caledonia in winter, our framework automatically counted individuals with a slight negative bias of −5.7% (−15 individuals over a total of 265 ground‐truth individuals) compared to a baseline method relying on the sum of detections that led to a 48 times overestimation. Our tracking‐by‐detection framework directly estimated the number of unique individuals per flight, leading to a mean seasonal index of abundance of 42 individuals (±28 SD) in the surveyed area. Overall, 35 individuals were missed across flights, representing 13.2% of all ground‐truth individuals. Among missed individuals, two‐thirds were calves superposed to their mothers, stressing the challenge of occlusion when counting group‐living animals. Despite being applied to entire videos with very low prevalence of dugongs, our framework generated just 8% of false positives (20 instances) across all flights, mostly due to misidentifications with other species. Illustrated through the case of dugong, our tracking‐by‐detection framework more generally provides a scalable method to support digital population surveys, benefitting long‐term monitoring and conservation programs of vulnerable megafauna.
Abstract Benthic organisms, such as sea cucumbers, play a crucial role in delivering essential ecosystem services, including nutrient recycling through feeding, excretion and bioturbation processes. Considering recent harvest activities in subtidal zones worldwide, it is imperative to define rapid and cost‐effective solutions to improve monitoring tools. Large‐scale Structure from Motion (SfM) photogrammetric mapping was conducted along Mediterranean infralittoral sea bottoms to provide training and testing datasets for developing a new object detection model. Several high‐resolution (0.5 cm/pixel) orthophoto mosaics were used to train a new object detection model based on the YOLOv11 architecture. Subsequently, georeferenced imagery was used to assess the accuracy of manual counts compared to the outcomes produced by the model. The model exhibited strong performance on the validation set, achieving a precision of 83% and a recall of 88%. The implementation of the open‐source GIS model through the Deepness plugin in QGIS software enabled the rapid deployment of the model (15.9 ± 1.35 min) over large (5500 m2) orthophoto mosaics, revealing the effects of habitat and seasonality on the detection results. The length of the diagonal of the bounding boxes used as a proxy for the body length of sea cucumbers, combined with a Gaussian kernel density estimator analysis, revealed comparable results between the estimated and observed values of sea cucumbers, confirming the validity of the proposed method for extracting biologically relevant information for natural population monitoring. This research presented a new and scalable workflow for identifying small benthic organisms through underwater photogrammetric‐based imagery. The findings underscore the promise of using cost‐effective and open‐source object detection tools to provide a valuable tool for ecological assessments of natural populations. Data from automatic detection systems represent a promising non‐invasive approach for directly analysing species distribution patterns based solely on image data, marking a significant improvement in ecological monitoring techniques.
Wetlands, especially peatlands, are fundamental to the functioning of the Brazilian Cerrado. They play essential roles in hydrological and carbon cycling, biodiversity conservation, and the support of local livelihoods. However, they remain poorly mapped and are not regularly monitored. Ongoing large-scale land use change combined with the climate crisis is intensifying the drying and degradation of these wetlands, highlighting the need for specific wetland assessments to inform adaptive conservation and management strategies. To extend the thematic details of current land cover products, we present a new approach to map swamp savanna and gallery forest, two ecologically distinct and carbon-rich wetland types in the Cerrado. We tested combinations of Sentinel-2-based spectral-temporal moisture metrics and Sentinel-1 polarized metrics from different time periods of a year using a U-Net architecture for image classification. Class-wise F1 scores reached 0.91 and 0.89, respectively, over the training area with homogeneous and consistent predictions. Adding Sentinel-1 VV and VH spectral-temporal metrics produced heterogeneous results, partly improving models. The segmentation models were applied with high accuracies to a transfer region of comparable environmental characteristics (F1: 0.85-0.90) and with lower accuracies to an environmentally differing region (F1: 0.67-0.73). From a methods perspective, the results suggest that upscaling to the entire Cerrado with such a model is feasible, despite the need for regional retraining for better model robustness.
Biocrusts are soil-surface assemblages of cryptogams, microbial communities, and soil particles, constituting the "living skin" of drylands. They are critical for global carbon and nitrogen cycles but are endangered by climate change. Safeguarding these critical components and their functions requires accurately measuring biocrust distribution and traits. Biocrust key functional indicators, such as pigments, provide information about their developmental stage, biomass, and functional state, but their remote quantification across spatial scales remains underdeveloped. We evaluated the potential of local-scale (tripod-acquired) multispectral imagery with machine learning to estimate pigments (scytonemin, chlorophyll a, carotenoids), soil organic carbon (SOC), and nitrogen (N) in two contrasting deserts: the Chihuahuan Desert (NM) and the Colorado Plateau (UT). High-resolution imagery (< 1 mm) alongside pigments and nutrient concentrations was collected. At NM site, support vector regressions (SVR) using blue and red bands achieved high accuracy for scytonemin (R-2 = 0.93); red-edge band enhanced predictions for SOC and N (R-2 = 0.95 and 0.80, respectively). In UT, predictive accuracy was lower (R-2 < 0.6), likely due to scytonemin saturation. Scytonemin-based predictions of biocrust cover were less sensitive to moisture variability than chlorophyll a or carotenoids, suggesting scytonemin's superiority as a biocrust index compared to previously developed indices. The best-performing local-scale models at the NM site successfully scaled to landscape-scale UAS imagery, allowing the remote prediction of pigments and nutrients and capturing relative differences associated with biocrust successional stages but did not capture the exact values of the modeled parameters compared to local measurements. The approach shows promise for future integration with satellite imagery to expand biocrust trait mapping at broader scales, offering a valuable tool for monitoring these key soil ecosystems and their functional attributes across diverse landscapes.
ABSTRACT Camera traps have become a popular tool for monitoring a wide range of species, but determining the animals present in the footage can be incredibly time‐consuming and remains a significant challenge for many projects. Crowdsourcing this task through the engagement of members of the public as citizen scientists can be a solution to this problem. There is good evidence that high confidence in classification accuracy can be attained from citizen science species classifications, but whether citizen scientists can also perform more difficult classification tasks, such as identification of individuals within a species, has received far less attention. Here, we used European pine marten Martes martes in the Forest of Dean, UK, as a case study to assess citizen scientist ability to identify different individuals from photo and video camera trap footage. Agreement between expert and citizen scientist individual ID classifications was variable, but there is potential for improving citizen science classification quality through optimising study design. Citizen scientists were more likely to classify pine marten footage correctly to species level and to attempt an individual ID when shown video footage rather than photos. Individual ID accuracy was also improved by the proximity of bait to a camera trap. Reassuringly, citizen scientists' assessment of their own confidence was positively associated with the accuracy of individual ID classification. This study highlights the potential for citizen science to contribute to more than just species classification in camera trapping projects but also emphasises the need for appropriate study design to ensure data quality.
Canopy cover modulates the subcanopy light regime in forests-variations in canopy cover alter both, the amount and spatial distribution of light, representing the source of energy for plants. Because plant species differ regarding their adaptations to light, canopy-driven light regimes play a key role in structuring understory composition and diversity. Here, we evaluated if Airborne Laser Scanning (ALS) and Sentinel-2 can map subcanopy light regimes across three study sites (n = 244 plot locations) along the German Alps. We then examined how understory vascular plant species richness varies with subcanopy light regimes. We found light availability increases linearly (Adj. R 2 0.84/RMSE 0.09) with ALS-derived gap proportion (i.e., inverse of canopy cover), whereas light variability (Adj. R 2 0.46/RMSE 0.05) follows a hump-shaped pattern. The relationship between the subcanopy light regime and vascular plant species richness in the understory (Adj. R 2 0.39/RMSE 12.77), was equally well described by ALS-derived gap proportion (Adj. R 2 0.40/RMSE 12.70). Canopy cover described by NDWI from Sentinel-2 showed similar relationships and could be used to characterize light availability (Adj. R 2 0.43/RMSE 0.18), light variability (Adj. R 2 0.21/RMSE 0.06), and understory vascular plant species richness (Adj. R 2 0.31/RMSE 13.70), albeit with lower variance explained. Understory vascular plant species richness increased with higher light availability and variability, indicating that canopy cover heterogeneity is an important driver of biodiversity next to overall canopy cover. Using spaceborne sensor systems might enable repeatable, large-scale mapping of canopy cover and thus also understory diversity, as our results suggest such sensors provide similar insights as gold-standard ALS data, which remain scarce. Understanding the links between canopy cover, the subcanopy light regime, and understory diversity is therefore essential for predicting vegetation dynamics and supporting biodiversity-oriented forest management.
Gaining insight into population changes of species in remote areas of the world poses challenges, often resulting in opportunistic observations (convenience sampling) that limit scientific inference. Remote sensing resources, typically high-resolution optical imagery, have revolutionised population ecology especially in polar regions like Antarctica and now many species have been censused. However, estimates based on remote sensing often remain limited to convenience sampling, rather than allowing observation at times of year that provide more information about the breeding population. Here we introduce a proof of concept using high-resolution synthetic aperture radar (SAR) imagery to detect, enumerate and track behaviours of emperor penguins during the Antarctic winter. Using 25-30 cm Umbra SAR imagery, we gathered images at six colonies of emperor penguins during winter 2024 to determine phenology, to gain insight to the breeding population and advance a new method for breeding population observations. We found that emperor penguins are identifiable on fast ice and that SAR-based observations match concurrent field observations (at Atka Bay). Total huddle area correlated with average colony size at emperor penguin colonies, and we found that larger colonies tended to use more space on the fast ice than smaller colonies. We demonstrate that the breeding population of emperor penguins during winter can be estimated using a combination of SAR imagery with phenological-behavioural models. Given that models suggest emperor penguins may be quasi-extinct before 2100, our work provides an important next step in understanding the approximate size and phenology of the breeding population, information that would be required for conservation of the species and for area-based conservation such as evaluation of the Ross Sea Marine Protected Area (MPA), where similar to 33% of emperor penguins exist.
Hair snares, consisting of strands of barbed wire that collect tufts of fur, have long been used as a noninvasive sampling technique for DNA-based monitoring of different species of bears. However, to prevent cross-contamination by other animals or different bear individuals, frequent visits by technicians are required to collect samples and clean snares, a highly demanding process in terms of time and resources that hampers scalability. In the Catalan Pyrenees, only 24.6% of the visiting efforts come out positive, highlighting the need for automated, scalable solutions that can operate independently of cellular networks or human intervention. This study proposes a novel methodology to automatically infer hair snare interactions by detecting bipedalism in images acquired by camera traps. This is implemented by using a state-of-the-art deep learning model for pose estimation, combined with a multilayer perceptron (MLP) to classify bipedalism or quadrupedy based on predicted keypoints. Due to the lack of annotated bear pose datasets, a custom dataset of 2373 images collected in the Catalan Pyrenees was manually annotated with 15 anatomical keypoints. Using YOLOv11, the trained pose estimation model achieved a keypoint inference precision of 93.2%, while the MLP reached an accuracy of 96.1% in distinguishing bipedalism from quadrupedy. Finally, to enable inference on the edge, several verification strategies are proposed using only the pose estimation output, including bear geometrical illustrations derived from the inferred keypoints. This innovative lightweight solution, designed for satellite-based transmission, enables scalable deployment in low-connectivity and remote areas, significantly reducing the need for frequent manual inspections.
Effective monitoring of cetacean occurrence is essential, as their populations face multiple anthropogenic threats, including shipping strikes, noise pollution, fisheries bycatch and habitat degradation. Passive acoustic monitoring (PAM) is a reliable non-invasive method for detecting cetaceans underwater, yet its wider adoption has been constrained by the availability and cost of equipment. Recent advances in cost-effective, user-friendly devices offer new opportunities for expanding acoustic data collection. In this study, low-cost acoustic recorders and underwater cameras were deployed across a range of fixed and mobile platforms in the Western Mediterranean, including fishing gear (drifting pelagic longlines and trammel nets), anchored Fish Aggregating Devices (aFADs) and boat-based observational surveys. These deployments demonstrate the practical potential for opportunistic cetacean presence data collection through collaboration with non-academic stakeholders, including the fishing sector and non-profit organisations. We report visually confirmed acoustic encounters of the sperm whale (Physeter macrocephalus) and four dolphin species (family Delphinidae), demonstrating that affordable PAM recorders can be used to capture cetacean vocalisations such as echolocation clicks and whistles. Despite certain limitations, these instruments can enhance the accessibility of PAM to a wide range of users, facilitating cetacean presence data collection and providing insights into interactions with human pressures, particularly fisheries. Moreover, low-cost devices demonstrate strong potential for integration into citizen science initiatives, offering a scalable solution to reduce operational costs and allow underwater soundscape data collection across large spatial scales. As PAM gains increasing recognition within ocean observing frameworks, our study underscores the value of low-cost PAM technologies in supporting efforts to document cetacean occurrence while fostering more inclusive approaches to marine biodiversity monitoring.
Bat migration is an ecologically important yet poorly understood phenomenon. This is in part because monitoring these migrations is challenging, due to bats' nocturnal behaviors and their sometimes high-altitude migratory flights. This study presents the first radar-based examination of multi-annual migratory bat phenology in Europe, utilizing vertical-looking radar data collected on the Swabian Plateau in Germany between September 2019 and December 2022. Bat activity was consistently low in winter and increased gradually from March onwards to a peak between July and September. Across all years, pre-maternity migration began between late February and mid-March, while post-maternity migration ended between late October and mid-November. We estimated peak radar-based migration traffic rates between 1159 and 2473 bats per km, with the highest peak recorded on 4 July 2022. Correlations between radar-derived nightly bat numbers and simultaneously acquired acoustic recordings ranged from 0.47 to 0.70 for the pre-maternity season, and from 0.14 to 0.71 during post-maternity migration. Both monitoring techniques showed peak bat activity during the summer, with smaller surges in September and October. The radar, however, detected significantly more bats overall. These findings showcase how vertical-looking radars can be used to quantify and characterize seasonal variability in high-altitude bat movements. Through strategic future radar deployments and the analysis of available historical datasets, our current understanding of migratory bat seasonality, routes, and intensity could increase drastically, and underpin the development of effective protocols for biodiversity conservation. Die Migration der Flederm & auml;use ist ein & ouml;kologisch wichtiges, aber noch wenig verstandenes Ph & auml;nomen. Dies liegt zum Teil daran, dass die & Uuml;berwachung der Flugbewegungen schwierig ist, da Flederm & auml;use nachtaktiv und h & auml;ufig in gro ss er H & ouml;he unterwegs sind. Diese Studie pr & auml;sentiert die erste radarbasierteUntersuchung von mehrj & auml;hrig erfassten ph & auml;nologischen Mustern ziehender Flederm & auml;use in Europa. Die Daten wurden mit einem vertikal ausgerichteten Radarger & auml;t erhoben, das zwischen September 2019 und Dezember 2022 in Deutschland auf der Schw & auml;bischen Alb in Betrieb war. Die Fledermausaktivit & auml;t war im Winter durchwegs gering und nahm ab M & auml;rz allm & auml;hlich zu, mit einem H & ouml;hepunkt zwischen Juli und September. In allen Jahren begann die pr & auml;-maternale Wanderung zwischen Ende Februar und Mitte M & auml;rz, w & auml;hrend die post-maternale Wanderung zwischen Ende Oktober und Mitte November endete. Wir ermittelten n & auml;chtliche Fledermausaktivit & auml;tspeaks zwischen 1159 und 2473 Flederm & auml;usen pro Kilometer, wobei der h & ouml;chste Wert am 4. Juli 2022 registriert wurde. Korrelationen zwischen radarerfassten n & auml;chtlichen Fledermauszahlen und gleichzeitig aufgezeichneten akustischen Daten lagen in der pr & auml;-maternalen Saison zwischen 0.47 und 0.70, und w & auml;hrend der post-maternalen Wanderung zwischen 0.14 und 0.71. Beide & Uuml;berwachungstechniken registrierten im Sommer die h & ouml;chste Fledermausaktivit & auml;t, mit kleineren Anstiegen im September und Oktober. Das Radar detektierte jedoch insgesamt deutlich mehr Flederm & auml;use. Diese Ergebnisse zeigen, wie vertikal ausgerichtete Radare verwendet werden k & ouml;nnen, um die saisonale Variabilit & auml;t von Fledermausbewegungen in gro ss er H & ouml;he zu quantifizieren und zu charakterisieren. Durch eine strategische Platzierung von Radarger & auml;ten und die Analyse verf & uuml;gbarer historischer Datens & auml;tze k & ouml;nnte unser derzeitiges Verst & auml;ndnis der Saisonalit & auml;t, des Verlaufs der Zugwege und der Anzahlen wandernder Flederm & auml;use erheblich erweitert werden, sowie die Entwicklung wirksamer Artenschutzprotokolle unterst & uuml;tzen.
The plateau zokor (Eospalax baileyi) is both an ecosystem engineer and a major rodent pest in the alpine meadow of the Qinghai-Tibet Plateau. Through its burrowing and soil-turnover activities, the species enhances soil aeration, promotes nutrient cycling, and increases microsite heterogeneity that can facilitate plant regeneration. Yet, its extensive mound building also reduces forage availability, alters vegetation structure, and accelerates soil erosion. These contrasting ecological roles make the abundance and positions of mounds key indicators for assessing rodent impacts and guiding grassland management. However, obtaining accurate and spatially explicit information on mound numbers and distribution remains difficult because traditional manual surveys are labor-intensive, spatially limited, and unsuitable for large-scale monitoring. UAV-based imagery provides new opportunities for automated detection, but the small size of mounds, their dense aggregation, and the heterogeneous background of alpine meadows pose substantial challenges for conventional algorithms. To address these limitations, we developed PZM-YOLO, an improved small-object detection model based on YOLOv5s that enhances feature extraction and localization accuracy for zokor mounds in complex UAV imagery. We compiled a multi-season UAV mound-image dataset and evaluated the model against several mainstream object-detection networks. PZM-YOLO achieved a precision of 91.1%, recall of 69.3%, AP50 of 76.1%, and an F1-score of 78.7%, outperforming the baseline YOLOv5s. Independent validation using 100 manually annotated UAV images further confirmed its reliability. The proposed framework enables automatic detection of plateau zokor mounds and provides quantitative information on mound abundance and mound position. This framework also offers a transferable approach for detecting indicators of subterranean rodents, such as mounds and holes, supporting ecological impact assessment and grassland restoration planning.
Crustose coralline algae (CCA) are key reef-building organisms, yet their fine structures make traditional visual surveys time-consuming and limit large-scale monitoring. This study evaluates the potential of unmanned aerial vehicles (UAVs) for monitoring CCA coverage on intertidal reefs along the Taoyuan coast, Taiwan. Three experiments conducted between 2022 and 2024 were designed to assess the effects of flight altitude, sensor type, classification approach, and the feasibility of large-scale surveys. A supervised Random Forest classifier was applied using five color indices (R-G, G-B, ExG, ExGR, and NGRDI) to evaluate their effectiveness and identify the optimal combination for CCA classification. Results indicate that classification accuracy was primarily governed by image resolution. Multispectral imagery, under typical UAV configurations, was insufficient to resolve millimeter-scale CCA features, whereas low-altitude visible-light imagery (< 2 m above ground level) provided sufficient detail for reliable identification. Although per-pixel classification accuracy remained moderate (Kappa > 0.4), the derived CCA coverage estimates showed strong agreement with reference observations (r > 0.87), indicating that low-altitude visible-light UAV imagery provides a robust basis for large-scale ecological assessments. Overall, UAV-based surveys combined with automated classification provide an efficient and scalable framework for mapping and monitoring CCA and reef habitats, with potential extension to other taxa.
Ecosystems worldwide are undergoing rapid degradation, yet effective monitoring remains costly and time-consuming. The proliferation of open-access imagery from satellites, Google Earth, and citizen-science platforms offers unprecedented opportunities to improve ecological monitoring, yet, their potential for species detection and demographic tracking remains underexplored. Here, we demonstrate a cost-effective approach to retrospective image analysis by combining current ground truth data with historical Google Earth RGB imagery to extract long-term demographic information. We apply this approach to two invasive plant taxa with contrasting growth forms in Mediterranean ecosystems. First, we use deep learning to detect individuals of prickly pear (Opuntia spp.) across diverse habitats and image resolutions, and reconstruct 10 years of spatially explicit recruitment rates along a climatic gradient. Second, we quantify nearly 20 years of growth dynamics for the clonal invader Carpobrotus spp. in two contrasting environments. Our object detection model achieves 60%-80% accuracy in identifying Opuntia individuals, with performance enhanced by colour consistency and contrast. While detection is limited for individuals <4 m(2), the approach captures similar to 80% of the population. For Carpobrotus spp., area-based analysis mapped >7,900 m(2) across two sites and revealed a mean genet expansion of 12.95 +/- 5.32 m(2) yr.(-1). Beyond detection, time-series analysis of Google Earth imagery allows the estimation of recruitment, growth rates, climatic sensitivity, population structure, size-age relationships, and recruitment hotspots. With image series spanning a decade for Spain, Greece, and the UK, and two decades for Portugal, we provide spatially explicit demographic reconstructions at unprecedented scales. By harnessing publicly available imagery, our pipeline expands the capacity for long-term, large-scale demographic monitoring. Although demonstrated here with invasive plants, the approach is broadly applicable across taxa and ecosystems. Retrospective image analysis has the potential to accelerate conservation, guide restoration, and support robust ecological forecasting in the Anthropocene.
Autonomous remote-sensing technologies are increasingly contributing to biodiversity monitoring by enabling scalable, repeatable, and minimally invasive data collection. We present a ground-based robotic remote-sensing framework that integrates artificial intelligence and standardized quality assurance to support the derivation of decision-ready ecological indicators. Using European coastal dunes as a case study, we deployed an AI-enabled quadruped robot equipped with near-ground imaging sensors to monitor the host-herbivore interaction between Pancratium maritimum and Brithys crini. In this citizen-to-robot pipeline, expert-verified citizen-science imagery was used to train lightweight detection models for on-board inference and higher-capacity models for offline auditing, ensuring reproducibility and transparency across missions. Field trials demonstrated that the system achieved consistent image quality, accurate detections, and low-disturbance operation under natural conditions, capturing spatially explicit evidence of herbivory and host condition. By coupling standardized protocols with robotic autonomy, this approach implements a proximal remote-sensing layer that complements aerial and satellite observations. The workflow is designed to support transferable quantification of species interactions and habitat condition across sites and seasons, contributing to the integration of robotics and ecological remote sensing for biodiversity assessment and conservation management.
Abstract Estimating tree life histories and population dynamics is key to predicting how forests respond to climate change and disturbance. However, linking individual tree trajectories to whole‐forest outcomes (e.g. structural, compositional, and functional health) remains challenging. Stage‐structured demographic models offer a promising solution, but they typically require extensive field data on individual‐level vital rates (e.g. survival and growth), limiting their application at scale. Here, we demonstrate an approach that integrates repeat airborne lidar data with a structured demographic model (an integral projection model, IPM) to examine forest‐wide demography in response to environmental drivers. Using Australia's Great Western Woodlands as a case study, we model the survival, growth, and life expectancy of ~40 000 eucalypt trees over a decade. Vital rates were modelled using height for small trees and crown area for large trees, reflecting a shift in growth strategy with size. Our results indicate distinct responses of small and large trees to proxies for competition and soil moisture (local canopy density and topographic wetness index, respectively). A reduction in topographic wetness index—reflecting drier conditions—led to lower life expectancy, particularly for larger trees, which may be more vulnerable to drought. This framework enables demographic analysis at scale, using widely available lidar data, offering a scalable tool for forest monitoring, modelling, and management. We identify three priorities for broader application, including (1) mixed species stands and multilayered canopies, (2) full life cycle modelling including reproduction and early life stages, and (3) long‐term or comparative studies using high‐quality repeat lidar. By combining remote sensing data with detailed insights from field‐based studies, our study provides a scalable approach for guiding forest management and conservation decisions.
