Incorporating explicit spatial heterogeneity in fuel characterization will reduce uncertainty in fire modelling, especially in Mediterranean forests which are characterized by complex structures. We provide a reproducible workflow to integrate bulk density estimates derived from terrestrial laser scanner (TLS) data into Fire Dynamics Simulator (FDS), and evaluate fire behavior sensitivity to wind speed and fuel moisture content (FMC) in two Mediterranean forest plots. The methodology incorporates the 3D fuel distribution into FDS derived from point clouds. 210 simulations were conducted by modifying wind speed and vegetation FMC. Results confirm the strong negative effect of FMC on fire behavior, the positive effect of wind on fire propagation, and evidencing the importance of their combined effect. Fire behavior sensitivity proved to be strongly scenario-dependent, demonstrating that structural heterogeneity plays a critical role with direct implications for assessing fire propagation probability under varying moisture and wind regimes in real forest stands.
UAV-based orchard monitoring can support yield estimation and precision management, yet reliable detection of small and densely distributed fruits remains difficult in complex field scenes. Challenges include limited object pixels, cluttered canopy backgrounds, frequent occlusion, illumination variability, and the need for real-time inference on resource-constrained edge devices. We develop Walnut Lightweight-Pruned YOLO (WLP-YOLO), a task-specific and deployment-oriented detector derived from YOLOv8 for UAV walnut detection under edge-computing constraints. WLP-YOLO combines lightweight feature extraction, efficient multi-scale fusion, hardware-friendly structured channel pruning, and target-device inference to balance small-object detection performance and computational efficiency. On the fixed dataset split, WLP-YOLO increased mAP@0.5 from 0.811 to 0.834 while reducing the model to 2.23 M parameters and 6.8 GFLOPs. Although mAP@0.5:0.95 remained comparable to YOLOv8n (0.323 versus 0.322), small-object AP@0.5 increased from 0.554 to 0.693. Further structured pruning reduced the computational cost to 2.7 GFLOPs, with only a 0.2 percentage-point decrease in mAP@0.5. On the Jetson Xavier NX, the pruned model achieved a per-image latency of 29.1 ms under TensorRT-FP16 inference, corresponding to approximately 34.4 FPS. These results support a practical accuracy–efficiency trade-off for UAV-based walnut monitoring, while transfer to other tasks requires task-specific retraining and cross-domain validation.
Effective conservation and long-term ecological monitoring of the giant panda (Ailuropoda melanoleuca), a flagship and nationally protected species, require quantitative behavioral characterization. Such analyses rely on accurate detection of anatomically consistent body landmarks that provide a structured representation of panda posture and movement. Most existing landmark detection and pose estimation methods have been developed for humans or controlled environments and therefore struggle with real-world giant panda monitoring. Low contrast, severe background clutter, and the fine-grained anatomical structures of giant pandas often result in unreliable landmark localization under natural field conditions. To address this issue, we propose Panda Pose Estimation (PandaPE), a domain-adaptive framework specifically designed for automated landmark detection in real-world giant panda monitoring. PandaPE incorporates three animal-centric modules tailored to the visual characteristics of giant pandas in infrared camera-trap images, enabling more robust landmark detection under challenging field conditions. We evaluated PandaPE on infrared camera-trap images of giant pandas collected in natural reserves, reflecting the realistic complexities encountered during long-term field monitoring. The experiments yielded average precision (AP), AP50, and AP75 scores of 47.0%, 90.0%, and 46.9%, respectively, demonstrating that the proposed method outperformed the baseline by 2.8%, 0.4%, and 9.5%, respectively. These results indicate that PandaPE enables anatomically consistent and stable landmark detection, providing a reliable foundation for automated posture analysis and downstream ecological and conservation studies.
Plant functional traits are important indicators of ecosystem functioning. Recent hyperspectral satellite missions such as EnMAP provide detailed spectral information that enables the monitoring of plant traits across large spatial scales. However, their coarse spatial resolution (30 m) affects validation efforts and applications in fragmented landscapes. Satellite scenes can be sharpened with image fusion techniques by combining hyperspectral data with finer resolution multispectral data. Yet, it remains unclear whether these fused images provide benefits for retrieving plant functional traits. In this study, we evaluated whether 3D-CNN-based sharpening of EnMAP with Sentinel-2 data preserves spectral fidelity, how it affects trait retrieval performance relative to original EnMAP, and whether it provides spatially more detailed plant trait maps in a riparian forest. The method was validated with field data collected in a riparian forest in Leipzig (Germany), including five key plant traits and fine-resolution airborne HySpex imagery. Trait retrieval was performed with a state-of-the-art deep learning model applied to original EnMAP data, sharpened EnMAP products, and EnMAP-like HySpex datasets at multiple resolutions. Results show that sharpening preserved spectral fidelity and increased the spatial detail of trait maps. Retrieval accuracy did not improve significantly, but paired error comparisons indicated no systematic degradation for chlorophyll, carotenoids, LMA, and EWT, while LAI showed increased error. Our findings indicate that 3D-CNN-based image fusion can support spatially more detailed EnMAP trait maps, but the products here should be interpreted primarily as representations of relative spatial trait variation rather than as accurate pixel-wise trait estimates. By preserving spectral fidelity while increasing spatial detail, the approach provides a pathway for future studies to extend the use of spaceborne hyperspectral data in monitoring ecosystem structure and function.
Seabirds are impacted worldwide by ongoing and emerging threats, such as habitat loss, unsustainable fishing and climate change. Within this context, models able to simulate and predict seabird behaviour depending on their environment would be valuable for decision making. In this study, we present the Spatial GAN (Generative Adversarial Network), able to simulate seabird foraging trajectories based on a spatial environmental condition. It is illustrated using GPS-tracking for Northern Gannets (Morus bassanus) from five different colonies. We demonstrate that the model simulates central-place foraging trajectories coherent with field data for each colony. The ability of the Spatial GAN to simulate trajectories with realistic global characteristics of length and duration consistently outperforms those of other existing models, including classic approaches such as HMM-informed Random Walks, and deep learning approaches such as VAE and the GAN without spatiality. The Spatial GAN also correctly simulates trajectories for birds from sites not included in the training dataset. While it may simulate erroneous trajectories on land when bathymetric configurations are new, the inclusion of only 10 to 20 GPS-trajectories from a new site greatly improves model simulations. As such, the architecture of the Spatial GAN we developed is a good starting point for a trajectory simulator using environmental maps as input data. The specificity of this approach compared to existing ones is its capacity to approximate the relationship between seabird movements and the environment structure by itself, without any explicit constraint.
Cetacean strandings provide early warning of emerging concerns regarding animal, ocean, and human health and allow insights into local cetacean population, density, distribution and mortality, and thus offer significant conservation value for the assessment of ecosystems. However stranding networks (programmes for monitoring and coordinating a ground response and investigation of local strandings) are infrequent to non-existent within minimally populated areas, coastal areas with limited economic resources, geographically remote areas, complex coastlines and areas of geopolitical unrest. Very high-resolution (VHR) satellite imagery offers the prospect of improving monitoring in these regions. While VHR optical satellite imagery is able to detect large baleen whale strandings (> 12 m), mass strandings (MSE, strandings of two or more animals excluding mother calf pairs) are predominantly of smaller-sized odontocetes (∼1–6 m). Detecting odontocetes is therefore crucial for VHR satellites to be useful for monitoring strandings globally. Here, we analyse MSEs involving large (∼11–18 m) and small odontocetes to; (1) investigate whether stranded animals can be detected and counted in VHR optical satellite imagery, (2) determine the minimum spatial resolution of VHR optical satellite imagery necessary for detecting and counting stranded animals, and (3) explore the utility of synthetic aperture radar (SAR) to detect and count stranded animals. We successfully detected and counted large and small stranded odontocetes in optical and SAR satellite imagery. With optical imagery, we found that 0.5 m spatial resolution is suitable for accurately counting large odontocetes (100% accuracy, one image) while 0.3 m resolution is most suitable for counting smaller odontocetes (82–100% accuracy, four images). With SAR, two or more observers detected a total of 108 small odontocetes with high certainty (‘likely’ or ‘definite’ detections) compared to 111 in a VHR optical satellite image collected five hours later. These positive results show the potential for stranding networks to expand their monitoring capacity both in terms of scale and speed of detection.
The Pacific coast of Canada is widely populated by bull kelp (Nereocystis sp.) and giant kelp (Macrocystis sp.), which create unique habitats that provide refugia and foraging opportunities for animals of all trophic levels. A recent decline in kelp forests has sparked concerns about the kelp populations' stability and the need for monitoring. Generally, optical spaceborne imagery is used to monitor kelp, but its application is limited by clouds. This study aimed to assess the correspondence between floating kelp extent determined using Synthetic Aperture RADAR (SAR) and that typically obtained using optical imagery. We evaluated three machine learning models: CATBoost, LightGBM, and XGBoost. The classification framework was based on annual changes in surface backscattering, since we observed elevated values from kelp in summertime, distinct from the surrounding waters. The XGBoost performed the best among the evaluated models, yielding a balanced accuracy of 0.76 compared with kelp extent derived from optical imagery. Our findings also revealed that data from July to September was best for mapping floating kelp using remote sensing since kelp remains consistently on the surface. This study proposes a new approach that allows observations regardless of cloud cover and captures changes in kelp forests throughout the year. This new capability opens many avenues for further research and provides new insights into the seasonal variability of the floating kelp canopy, often driven by responses to marine stressors, thus allowing managers to assess disturbance response.
The global transition to low carbon energy has increased demand for strategic salt lake minerals such as lithium and potassium, while intensifying the tension between resource development and the protection of fragile ecosystems. Conventional ecological assessments rely largely on static background indicators and geometric buffers, which limit their ability to represent disturbance pathways across environmental media, such as salt dust dispersion and water and salt migration, and to account for climatic covariation in ecological change. Taking the Qinghai–Tibet Plateau as the study area, this study developed a regional ecological stress screening framework that integrates hydrological connectivity constraints, correction using seasonal wind fields, adjustment for climatic background conditions using multiscale geographically weighted regression, and regional extrapolation using a random forest model. The modeled disturbance fields exhibited pronounced differences among environmental media in spatial extent, transport direction, and seasonal variation. The hydrologically constrained soil and water disturbance field covered 66,745 km2, including 13,349 km2 classified as high stress areas. Analysis of seasonal wind fields identified spring as the dominant season for potential atmospheric transport of salt dust, accounting for 48.27% of the annual transport potential, with transport mainly associated with westerly and west northwesterly winds. After statistically accounting for the selected climatic covariates, development disturbance was generally associated with adverse changes in ecological health. This negative association, as estimated by the model, tended to be more pronounced in areas with a greater number of heavy precipitation days during the warm season. Within the influence zones of existing developed mining areas, higher modeled risk classes were generally accompanied by higher deterioration proportions for several ecological indicators, providing partial evidence of consistency between ecological responses and the local risk ranking. Extrapolation across the plateau using the random forest model indicated relatively higher modeled suitability in the southern Tibetan mountains, the Nagqu hilly uplands, and the broad valleys of the southern Qinghai Plateau. Relatively lower modeled suitability occurred in the lake basins of the Qiangtang Plateau and along both flanks of the Kunlun Mountains. The Qaidam Basin showed moderate modeled suitability and pronounced spatial heterogeneity. These results provide regional screening evidence for prioritizing subsequent field investigations, analyses using process models, and environmental assessments at individual sites.
Antarctica is highly vulnerable to climate change, with rapid warming, cryospheric loss, and atmospheric changes increasingly affecting ecosystem stability and species persistence. This study integrates habitat suitability modelling with long-term environmental pressure analysis to identify spatial priorities for conservation across the Antarctic region. Habitat suitability indices (HSI) were developed for Antarctic avian and mammalian genera using four machine learning algorithms (Random Forest, CART, Gradient Boosting, and Maxent) and multisource satellite data. Long-term environmental pressure was quantified through pixel-wise trend analysis of surface temperature, snow and ice indicators, vegetation proxies, moisture indices, and ozone concentration derived from MODIS and Sentinel-5P observations. The two indices were normalized and integrated into a conservation priority framework, resulting in 16 combined suitability–pressure classes. Results indicate that most of Antarctica is characterized by low habitat suitability under moderate to high environmental pressure, while a limited proportion of coastal regions exhibits high habitat suitability despite increasing environmental stress. Approximately 1% of the Antarctic land area was identified as high-suitability habitat experiencing high environmental pressure, representing priority zones for targeted conservation intervention. In contrast, the Mawson Coast consistently exhibits high habitat suitability under relatively low environmental pressure, highlighting its potential role as a conservation refuge. By explicitly linking habitat suitability with long-term environmental trends, this study provides a spatially explicit framework to support conservation planning under climate change in data-sparse polar environments.
Species distribution models (SDMs) are useful tools to predict species’ potential distributions in under-sampled, novel environments. However, available data often fails to represent the full range of viable environments for various reasons, including sampling biases, biogeographic barriers, and the availability of potentially suitable habitats. Therefore, many SDMs are likely based on truncated niches and will extrapolate in unsampled environments, ultimately over- or under-estimating species’ distributions, especially under climate change. To identify optimal modelling decisions to maximize SDM extrapolation performance, we generated 1007 virtual species, then subsampled virtual occurrences from the northernmost 80% virtual presences to deliberately generate a dataset representing truncated niche. This dataset was used to train SDMs which extrapolate in hot, non-analog environments as in climate change assessments. This workflow allows us to test how different SDM algorithms, SDM ensemble methods, and variable selection strategies improve SDM extrapolation performance. Niche truncation, caused by the deletion of southernmost 20% occurrence data, increased inaccurate predictions by more than 60%. SDMs may overpredict species distributions when extrapolating in unsampled environments, distorting our estimation of climate change response and invasion risk. When data representing full range of species’ viable environments is unavailable, decisions on algorithm, variable, and ensemble methods can improve SDM (extrapolation) performance across time and space to estimate climate change consequences and manage invasion risks. We suggest including all potentially relevant variables, reporting variations in ensemble model predictions caused by different ensemble methods, and using different sets of SDM algorithms to minimize either overprediction or underprediction depending on modelling goals.
Coastal ecosystems are biodiversity hotspots, but the increasing anthropogenic pressure makes them vulnerable. The resulting ecosystem degradation has led to a rise of the scientific and local communities' concern, calling for monitoring systems. In this line, the adoption of remote sensing technologies is increasing as their costs decrease, especially Unmanned Aerial Vehicles (UAVs) in aerial imaging. However, its deep-learning based image analysis remains limited in coastal ecology, particularly for monitoring intertidal population dynamics. Existing methods often rely on labor-intensive in situ sampling, restricting spatial and/or temporal scales. In this work, we develop a UAV-deep learning pipeline for monitoring small, cryptic intertidal species, overcoming limitations of traditional methods. We use the faecal casts of Arenicola marina as a case study to demonstrate the methodology.The workflow combines two You-Only-Look-Once (YOLO) models for object detection and instance segmentation to estimate population abundance and size structure respectively. The models are trained on UAV high-resolution images annotated for faecal casts, recruit casts, and cast tubes. Our detection model achieved a precision and recall over 80%. For the instance segmentation model, a filtering process (detection filter prior to segmentation) was included to reduce false positive ratio. The instance segmentation model measured over 28,700 cast tubes in around 8600 faecal casts, with a strong correlation between manual and automated measurements. Compared to traditional methodologies, the UAV-deep learning one expanded the surveyed area by a factor of 2.5, increased the number of samples by a factor of 90, and notably reduced the sampling effort.
Fishery-independent video surveys are increasingly used to assess scallop populations due to their minimal environmental impact and high precision compared to traditional dredge methods. However, manual analysis of video footage is time-consuming and resource-intensive, limiting the efficiency and scale of survey analysis and delaying fishery management decision making. This study developed and evaluated deep learning models to automate scallop detection, counting, and sizing from underwater video surveys in Great Oyster Bay, Tasmania. We trained YOLOv8 object detection models on annotated datasets from 2021 and 2022 towed video surveys to identify live scallops, dead scallops, and other bivalves. Detections were linked across frames with BoT-SORT tracking so each individual was counted once. Separate models were trained to detect laser scaling points for calibration of size measurement. YOLOv8 models achieved scallop detection performance of mAP50 0.25–0.45 across all configurations, with live scallop detection reaching mAP50 0.54–0.75, while laser detection models achieved mAP50–95 > 0.81. Counting performance showed contrasting patterns between survey years: substantial undercounting in 2021 (19% of manual counts) due to limited training data, and closely matching counts in 2022 (100.4% of manual counts). Exploratory sizing analysis produced broadly similar size distributions (manual mean: 107.3 mm, SD: 18.1 mm; automated mean: 102.3 mm, SD: 23.3 mm), though a secondary cluster of small false positive detections inflated the sub-legal proportion in automated measurements. While this approach may not yet fully replace manual analysis, it demonstrates significant potential to transform scallop monitoring by enabling faster, scalable, and more consistent interpretation of underwater survey data.
Leaf inclination angle distribution (LIAD) is a key descriptor of canopy structure that governs light interception, evapotranspiration, rainfall capture, and biomass production. Accurate LIAD estimates support ecophysiological modeling, enhance crop phenotyping, and precision agriculture; yet LIAD remains among the least constrained parameters in land-surface models due to measurement challenges. We explored a transfer-learning framework for classifying LIAD into five classic types from leveled RGB images. The full model configuration uses an ImageNet-pretrained EfficientNet-B0 backbone with custom fully connected layers to capture leaf-orientation features. The last five backbone stages were fine-tuned while earlier stages remained frozen to preserve generic visual representations. To assess each component's contribution, we conducted a component-wise ablation over four elements: backbone fine-tuning, custom classifier head, image-level augmentation, and mixed-sample training. On the model-development held-out test set, the configurations performed comparably, with macro-F1 values ranging from 85.26% to 88.34%, indicating that no single component dominated performance. Mixed-sample training was the most beneficial component, as its removal caused the largest macro-F1 reduction of 2.02 percentage points. In contrast, disabling image-level augmentation slightly improved held-out test performance and produced the highest component-ablation result, with 88.16% accuracy and 88.34% macro-F1. On the independent validation dataset, the same reference configuration achieved 52.83% accuracy, outperforming two prior machine-learning approaches by 27.43 and 21.08 percentage points. These results indicate that transfer learning can support LIAD classification from leveled RGB images, while further validation across species, sites, and acquisition conditions is needed before broader operational use in canopy-structure assessment and ecological or precision-agriculture workflows.
Non-native species (NNS) can have significant ecological and economic impacts worldwide, making timely monitoring and effective communication essential. We used YouTube metadata to examine how NNS with first records in the Iberian Peninsula between 2006 and 2025 are represented on the video-sharing platform and how this discourse is thematically framed. Using the YouTube Data API, we built a multilingual, Iberia-validated corpus from video titles and descriptions, classifying keywords and expressions into four broad themes: Invasion, Detection, eCommerce, and Threat. We combined descriptive analyses of thematic proportions with multinomial and generalised additive models to assess variation across taxonomic groups, invasion status, introduction period, and time since first record. Audience response was assessed using an engagement index derived from video metrics. After manual validation, the dataset comprised 1895 videos from 1002 channels covering 94 species. Thematic composition varied in relation to first-record year, with decreased trade-related content after first records and invasion and detection themes becoming more frequent. Video production was highly decentralised across channels, with insects achieving the highest popularity. Several species with limited video content nevertheless showed high per-video engagement, indicating that online visibility is not determined solely by upload volume. These results show that YouTube metadata can complement traditional monitoring by relating ecological risk to societal attention and revealing how NNS become socially visible and framed in public discourse. The multilingual lexicon and workflow developed here support conservation culturomics, targeted communication, and potential integration of digital discourse indicators into horizon-scanning frameworks in biodiversity-rich regions under high introduction pressure.
Projections of the agricultural non-CO2 greenhouse gas (NCGHG) transfers under diverse regional development pathways offers vital insights for achieving both climate goals and responsible consumption. We calculated the agricultural NCGHG emissions in 30 Chinese provinces from 2007 to 2020, and examines the spatio-temporal dynamics of interregional emission transfers using an environmentally extended multi-regional input-output model. Combining this model with the STIRPAT model, we simulated future embodied NCGHG emissions within regional trade under multiple scenarios. The results revealed an upward trend in net interregional transfers of NCGHG emissions from 87.4 Mt. CO2eq in 2007 to 194.7 Mt. CO2eq in 2020. These flows were mainly aggregated in East and South China. East China remained the dominant inflow region, reaching 125.13 Mt. CO2eq in 2020, while Northwest and Northeast China became the largest net outflow regions. Scenario simulations show that future consumption-based agricultural NCGHG emissions were highest under an extensive development scenario and lowest under green development, with the mitigation gap between them reaching 143.1 Mt. CO2eq by 2050 and East China evidencing the largest reduction potential. Central, and Southwest China also faced heightened emission pressures, highlighting China's regional disparities in future mitigation potential. Crucially, the results revealed an interregional carbon leakage mechanism: net import regions satisfy their agricultural demands through trade, thereby externalizing the environmental costs of NCGHG emissions to net export regions Our findings provides essential scientific evidence to supporting the need for region-specific mitigation strategies for facilitating China's realization of its carbon peaking and carbon neutrality goals.
Rice paddy fields contribute approximately 10% of global agricultural methane (CH₄) emissions, representing a major source of greenhouse gases from human activities. Methane is released continuously through diffusion in water and plant-mediated transport, as well as through episodic ebullitive events characterized by sudden gas bursts. While continuous fluxes are relatively well understood, ebullition is difficult to predict because of its sporadic and non-linear nature, introducing significant uncertainty into emission estimates and carbon budgets.This study evaluates the potential of machine learning (ML) techniques to estimate the occurrence of ebullitive CH₄ events and quantify their fluxes, aiming to improve emission estimates and support mitigation strategies. Input data were obtained from field measurements carried out in rice paddies at the Riet Vell farm (Ebro Delta, Spain) using a closed-chamber system coupled to an infrared gas analyzer (LI-COR 7810). Environmental predictors, including CO₂ flux, treatment, photosynthetically active radiation (PAR), water column, relative humidity, air temperature, soil pH, soil conductivity, and soil temperature, were also recorded. Candidate ebullitive events were identified from the first derivative of the CH₄ concentration time series and subsequently confirmed by manual inspection. Bubble occurrence was predicted using Decision Trees, Naïve Bayes, Random Forests, Neural Networks, Logistic Regression and Support Vector Machines, whereas Linear Regression, Random Forest Regressor and Gradient Boosting were evaluated for ebullitive CH₄ flux prediction under a leakage-safe cross-validation framework.Random Forest achieved the best overall performance for bubble detection, whereas Linear Regression provided the highest predictive performance for ebullitive CH₄ flux quantification. CO₂ flux was consistently identified as the dominant predictor, together with treatment, soil conductivity and, to a lesser extent, water column. These findings demonstrate that machine learning can capture complex relationships between environmental variables and methane ebullition while providing a robust framework for analyzing high-frequency chamber measurements. The proposed approach may contribute to improving greenhouse gas inventories, supporting sustainable rice management, and informing methane mitigation strategies in flooded agroecosystems.
Accurate seasonal flood prediction is increasingly critical in semi-arid regions under climate change, yet traditional models often struggle to capture the nonlinear hydrological responses in diverse catchment types. This study develops and evaluates a hybrid machine learning model combining Support Vector Regression and XGBoost for predicting three-month average spring discharge as a proxy for seasonal flood potential. The dataset includes 41,447 daily discharge records spanning 66 years (1952–2017) from four hydrological stations in western Iran. The hybrid SVR-XGBoost model was compared against five benchmarks: SVR, Random Forest, Gradient Boosting, XGBoost, and Multilayer Perceptron (MLP). The SVR model performed best at three stations: Gawshan, Ghorbaghestan, and Pol-e Haji Abad. XGBoost excelled at the urban Taq-e Bostan station. MLP showed severe overfitting across all stations, producing negative R2 values. The hybrid SVR-XGBoost model outperformed the best individual model at every station. It achieved a mean R2 of 0.724 and an RMSE of 0.606, a 22.9% improvement in R2 and a 19.6% reduction in error relative to the best single models. The largest improvement was at the mountainous Gawshan station. The highest absolute performance was at Taq-e Bostan. Feature importance analysis showed distinct seasonal controls. March discharge was the dominant predictor in mountainous catchments. November discharge played the most significant role in plain stations. The hybrid SVR-XGBoost framework captures catchment-specific hydrological dynamics. The model shows promise as a framework for flood warning systems, with potential application in semi-arid regions such as Kermanshah Province. Further validation across diverse climatic and hydrological conditions is needed to confirm its generalizability.
Marine ecosystems are characterised by complex, dynamic, and non-linear interactions that challenge traditional ecological modelling approaches. Recent advances in computational methods, particularly Graph Theory (GT), Graph Neural Networks (GNNs), and Reinforcement Learning (RL), offer promising tools for analysing, predicting, and managing these systems. This study presents the first comprehensive systematic review that jointly examines the application and integration potential of GT, GNNs, and RL in marine ecology. Following a PRISMA-guided methodology, 122 studies published between 2000 and 2026 were analysed to evaluate methodological trends, application domains, theoretical rigour, and practical relevance.The results show that GT remains the most widely used and mature approach, particularly for analysing ecological connectivity, food webs, and conservation planning. In contrast, GNNs and RL are emerging methods, with GNNs primarily applied to prediction and classification tasks (e.g., sea surface temperature forecasting and species behaviour analysis), and RL focused on adaptive decision-making in dynamic and uncertain environments, such as fisheries management and autonomous monitoring systems. Despite their individual strengths, these methods are largely applied in isolation. Key challenges identified across the literature include data sparsity, limited interpretability, weak integration of ecological knowledge, and insufficient real-world validation, particularly for RL.To address these gaps, this review introduces a formal scoring framework to assess theoretical rigour across six dimensions: ecological relevance, theoretical foundations, uncertainty quantification, scalability, stakeholder utility, and interpretability. Furthermore, the study highlights significant geographical disparities in research contributions and emphasises the need for greater inclusion of biodiversity-rich regions in the Global South. Finally, we propose the development of integrated GT–GNN–RL frameworks as a promising direction for future research, enabling the modelling of adaptive behaviours within networked ecological systems and supporting explainable, data-driven decision-making for sustainable marine conservation.
Savanna ecosystems cover approximately 20% of the global land surface and support substantial woody biomass, yet their forest extent remains poorly quantified. Accurate forest extent maps are essential for ecosystem monitoring, carbon accounting, and biodiversity conservation. However, current methods are optimised for tall-tree forests and may be inaccurate in other biomes. This work evaluated a recently published dual-polarimetric decomposition method, applicable to SAR data with and without phase, for mapping forests in tropical savannas. The decomposition was applied to Sentinel-1, NovaSAR-1, and PALSAR-2 data, each stacked with Sentinel-2 multispectral imagery. Random forest and extreme gradient boosting classification algorithms were trained on canopy height models produced by drone-mounted lidar. The new decomposition significantly outperformed the dual-polarisation entropy-alpha method, and decomposition features ranked among the most important predictors. PALSAR-2 produced the highest single-sensor accuracies, while PALSAR-2 and Sentinel-1 stack combined with Sentinel-2 yielded the best overall results. The XGB classifier achieved a k-fold cross-validated weighted average F1 score of 77.7% and an F1 score of 86.6% for forest prediction. The total detected forest area of 1033 km2 vastly exceeds the estimate from Australia's National Forest Inventory for the same region, indicating a substantial underestimation of forest extent in existing national datasets. These methods represent a traceable path for forest extent estimates, with important implications for regulated biodiversity and carbon market policy.
