.
Long-term animal re-identification (ReID) must remain robust to gradual morphological evolution and seasonal appearance shifts. Although recent vision–language models provide strong pretrained visual representations, adapting them to longitudinal ecological settings remains challenging, particularly under identity and temporal distribution shifts. We present a parameter-efficient CLIP adaptation framework for animal ReID and introduce a continuous metadata-conditioning mechanism that incorporates numerical attributes directly into the prompt representation during training. While low-rank visual adaptation, prompt-based supervision, and cross-modal alignment provide the adaptation framework, the proposed metadata-conditioning strategy constitutes the primary methodological contribution. By preserving the continuous structure of numerical metadata rather than discretizing it into textual categories, the proposed approach enables smooth modulation of the embedding space during training while maintaining a purely visual inference pipeline. Experiments on a seven-year longitudinal fish dataset and multiple wildlife benchmarks demonstrate improved performance under closed-set, open-set, and time-aware evaluation protocols. The results demonstrate that continuous metadata conditioning improves robustness to longitudinal appearance variation and temporal distribution shifts, while parameter-efficient adaptation enables a purely visual inference pipeline without requiring metadata at test time. Code and evaluation splits can be found at: https://github.com/AnilOsmanTur/MetaPrompt-ReID.
The lack of regular monitoring actions associated with high costs and long-lasting taxonomic expertise prevents the environmental assessment and implementation of conservation strategies in many regions of the World Ocean, especially in the tropical Large Marine Ecosystems, like the Gulf of Guinea, a basin influenced by numerous anthropogenic threats associated with oil excavation, and at the same time, very scarcely studied. Our study aimed to compare the effectiveness of the standard monitoring protocols based on Van Veen grab samples, with video recordings obtained using a Remotely Operated Vehicle (ROV) along the coast of Ghana in the 250–1000 m depth gradient. We have applied a variety of multivariate methods and diversity measures to analyse responses of benthic communities to multiple environmental factors of natural and anthropogenic origin. Even though both gears collect completely different faunas and pointed at possible sampling bias associated with specific features of each sampling gear, the analysis allowed to detect the same most important disturbance agents, like barium and hydrocarbons. The ROV material allowed for relatively quick taxonomic analysis and provided a reliable dataset allowing to obtained results of efficiency and reliability comparable to grab samples that are more effort-consuming for the taxonomic analysis.
The CNES-CLS22 Mean Dynamic Topography (MDT; https://doi.org/10.24400/527896/A01-2023.003, Jousset, 2023) represents an incremental update to previous CNES-CLS solutions, combining altimetry, satellite gravity, and in situ observations (drifters, hydrography profiles, and HF radar data). The main improvement lies in the Arctic, where enhanced Mean Sea Surface (MSS) coverage eliminate artifacts present in CNES-CLS18 and enable a more physically consistent representation of circulation, including the Norwegian Atlantic Front Current along the Mohn Ridge. Globally, CNES-CLS22 remains close to CNES-CLS18, with modest improvements in validation against independent datasets: RMS differences in geostrophic velocities decrease by only ∼ 0.2 %–0.5 % at the global scale and the average variance reduction at the global scale compared to heights derived from profiles is ∼ 0.5 %. Though regional gains are significant in the Arctic and Nordic Seas. HF radar integration in the Mid-Atlantic Bight demonstrates progress but highlights persistent challenges in shelf regions dominated by ageostrophic processes. At very small scales (< 40 km), noise from in situ data may introduce unrealistic kinetic energy, underscoring the need for improved filtering. Overall, CNES-CLS22 consolidates previous advances and provides better representation of key circulation features, but further progress will require enhanced coastal observations and refined processing methods, particularly for high-latitude and shelf areas.
Deep learning (DL) is a powerful tool to extract ecological information from large image datasets efficiently and consistently. However, applying these methods remains challenging, due in part to the complexity of DL workflows and the dynamic nature of available tools. To address this, we created a practical guide and review, focused on one of the fundamental tasks in automated image analysis: image classification. Our approach integrates commonly used software and highlights key steps-from image acquisition to annotated, model-ready datasets, to training, evaluation and deployment. It is modular and supported by a flexible code base (in Python, with R alternatives where possible) and Graphical User Interfaces (GUIs), enabling adaptation to different models and ecological objectives. The goal is to empower ecologists to confidently incorporate computer vision into their research. We illustrate this approach, using an open-source ROV dataset from the Norwegian Sea, featuring deep-sea biotopes defined by multivariate clusters of depth, substrate type, and associated species. To balance accessibility for users alongside performance, we focused on CNN models from the Ultralytics ML Platform (YOLO V.8 and V.11), comparing the full suite of architectures that range in complexity and efficiency. Cross-validation revealed high overall performances and that larger, more complex models are not always superior, with YOLO V.8m best (accuracy and macro-averaged performance metrics = ~0.97-0.98). Notably, high performances were achieved despite labels being based on both visual and external environmental predictors, suggesting visual features alone were sufficient for classification in this dataset. We highlight that the decision to deploy a model must be made in light of the study's objectives, with domain-based reasoning and experience guiding every stage of implementation. This work offers a practical blueprint for implementing DL in ecological research, promoting broader adoption and supporting reproducibility and more efficient, standardised, and sustainable monitoring; in this case of deep-sea biotopes, which is essential for marine spatial planning.
Urbanization, characterized by the expansion of impervious surfaces, significantly alters watershed hydrology. During precipitation events, these surfaces generate urban runoff, a hotspot of microplastics (MPs) that pose potential threats to human health and ecosystems. Substantial field surveys have been undertaken to explore the dynamics of MPs in urban runoff within small catchments (typically below 1000 km2). Nevertheless, identifying potential sources of MPs in large regions remains a challenging task. In this study, we provide a polymer-type-specific exploration of MPs in a large metropolitan area located on a piedmont alluvial fan, spanning over 16 000 km2. A total of 20 759 MPs (size ranging from 20 to 500 μm), representing 11 polymer types, were identified using laser direct infrared (LDIR) chemical imaging spectroscopy. Different sampling types, including impervious surfaces, roof drainage, and soil slopes across both hilly and plain areas, were included in this study. Multivariate statistical analyses, including partial least squares path modeling, showed that the abundance of MPs was influenced by precipitation characteristics, topography, and degree of rurality. Furthermore, multiple lines of evidence from abundance, polymer type, oxidation, and fouling characteristics suggested contributions from both local mobilization of soil MPs and atmospheric deposition. Results from this study are encouraging for the source identification of MPs in large areas.
