Paul Tresson

Long-term monitoring of tropical forest dynamics is constrained by limited historical fielddata and the short temporal coverage of satellite observations. Here, we reconstructed sevendecades of land-cover change on Reunion island (South-West Indian Ocean) by combiningdeep-learning-based feature encoding with random forest classification of orthomosaics fromaerial photographs acquired in 1950 and 2022. The classifiers achieved high accuracy,particularly for forest detection across wide structural and compositional diversity. Between1950 and 2022, Reunion underwent major land-cover transformations, including a tenfoldexpansion of urban areas and a net 6% increase in forest cover (+160 km2), driven by forestgains (+380 km2) and forest losses (-230 km2). Forest gains and losses were concentrated atlow elevations: gains mainly resulted from secondary succession on abandoned agriculturaland low-vegetation lands, whereas losses were driven by conversion of existing forests tothese classes. Most stable forests occurred at mid- to high elevation, generally withinprotected areas. Forest dynamics led to reduced fragmentation by 2022, with fewer, larger,and more aggregated forest patches. Independent data showed that newly forested areaswere disproportionately invaded by exotic plants, especially on the dry leeward side, while aminority of new forests on the wetter windward side showed resilience under lower invasionpressure. Even stable forests were substantially invaded, indicating that biological invasionspose a pervasive risk regardless of forest history. Overall, this study highlights a powerful,scalable framework for leveraging historical aerial imagery for tropical forest ecology andconservation.

Monitoring tropical forest structure at landscape scale requires cost‐effective methods capable of bridging the gap between field inventories and satellite remote sensing products. However, the diversity of available individual tree crown (ITC) segmentation algorithms raises questions about the consistency and reliability of derived structural metrics—a critical issue for any ecological application relying on these outputs. We evaluated three ITC segmentation algorithms—Detectree2, SAM, and the hybrid Detectree2SAM (D2S)—applied to very high‐resolution RGB orthomosaics (5 cm resolution) acquired over ~200 ha of the Luki Biosphere Reserve (Democratic Republic of Congo) using a low‐cost drone. Algorithms were validated at the individual scale against a photointerpretation reference of 1882 manually delineated crowns and at the plot scale against a field inventory of 360 trees across 18 plots. At the individual scale, IoU‐based F1 scores ranged from 0.57 (D2S) to 0.67 (SAM), revealing clear precision–recall trade‐offs. At the plot scale, D2S provided the most accurate crown area estimates (RMSD = 26%–29%), while SAM best reproduced canopy density (RMSD = 22%). All three algorithms systematically overestimated aggregated metrics; a linear correction substantially reduced prediction errors and improved cross‐algorithm convergence, yielding consistent estimates—median crown area of the 20 largest trees (~148 m ² ), total crown area (~3850 m ² per plot), and canopy density (~80 ind/ha). Agreement maps revealed systematic spatial divergences, including edge artifacts, SAM's tendency to underestimate canopy density, and localized overestimation of crown areas in disturbed stands. Raw ITC outputs carry substantial systematic biases requiring explicit field‐calibrated correction before ecological use. We propose a three‐step operational pipeline—(i) plot‐scale field validation, (ii) per‐algorithm linear bias correction, and (iii) landscape‐scale aggregation—that reduces inter‐algorithm divergence and delivers consistent, ecologically interpretable structural estimates from low‐cost UAV imagery.

Le charançon noir Cosmopolites sordidus (Germar Col. Curculionidae) est le principal ravageur des bananiers. Afin de développer le contrôle biologique par conservation de C. sordidus, il est nécessaire d'identifier les espèces impliquées et de quantifier leurs impacts en termes de régulation. Les méthodes classiques pour estimer la régulation d'un ravageur par ses prédateurs sont souvent biaisées : limitées à certains groupes, avec des biais d'échantillonnage, sous-estimant certaines interactions. Afin de limiter ces biais, ce travail vise à utiliser une approche d'imagerie in situ et d'analyse automatique pour mieux comprendre les réseaux trophiques liés à C. sordidus.Des expériences de proies sentinelles sous des caméras puis l'analyse des séquences d'images par des réseaux de neurones convolutifs se sont avérées efficaces pour déterminer les espèces présentes et leurs interactions (86,6 % de précision et 88,9 % de rappel pour la détection de 21 espèces différentes). Cette méthode rend également possible l'observation de la dynamique temporelle des évènements de prédation, et des interactions non-trophiques comme la coopération entre individus d'une même espèce. Cette méthode est ensuite améliorée en intégrant une étape de classification hiérarchique afin d'améliorer la robustesse des analyses face à des jeux de données biologiques contraignants.Sur l'île de la Réunion, des expériences de proies sentinelles vivantes (charançons adultes et œufs) ont été conduites sur cinq parcelles. Trois mille heures d'observation ont permis la capture de 375 000 images. L'analyse automatique des images (par soustraction de fond) a permis d'identifier les prédateurs et de quantifier la prédation des charançons adultes (33,6 %) et des œufs (58,7 %). De manière inattendue, la régulation des adultes de charançons a été effectuée par des vertébrés (Musaraignes Suncus murinus, Souris Mus musculus, lézards Calotes versicolor et crapauds Sclerophrys gutturalis). Jusqu'ici, le rôle des vertébrés dans la régulation du charançon du bananier semble donc avoir été sous-estimé ; la recherche d'ennemis naturels du charançon s'était jusqu'alors concentrée sur des arthropodes (fourmis, dermaptères, coléoptères, araignées).Ces résultats permettent de revisiter l'aménagement des parcelles afin de favoriser le contrôle du charançon (habitats favorables aux vertébrés et à leur prédation). De manière plus générale, ces travaux suggèrent que l'importance des vertébrés dans le contrôle biologique pourrait être sous-estimé. Ces résultats confirment le potentiel de l'analyse d'images pour des études en écologie.

Quantitative wood anatomy requires precise measurement of wood cells. This step is often laborious and limiting for further analysis. We introduce Samwood, a tool based on the zero-shot Segment-Anything Model to easily segment cells on microscopic images without the need for a training dataset. The reconstruction of cell files then allows for the analysis of wood along growth orientation and precise measurement of anatomical properties of the wood, such as lumen areas. We tested our pipeline on an example dataset of fossil woods featuring deformation, heterogeneous preservation, and frequent artefacts, to assess the robustness of our approach. The model achieves a precision of 0.78 and a recall of 0.80, often producing segmentation of better quality and more consistent than a human operator. This approach substantially reduces analysis time, minimizes operator bias, and provides a robust and extensible framework for large-scale anatomical studies The complete code pipeline is available at https://github.com/umr-amap/samwood . GRAPHICAL

I've appreciated reading this paper on reference-free size estimation in ants. The integration of multiple tools and techniques to overcome the technical challenges displayed is very interesting, and the underlying science seems to be sound. I do h...

Understanding of ecological interactions is necessary for the application of biological control. Banana is the second most produced fruit worldwide and the banana weevil (Cosmopolites sordidus) is the most important pest of banana and plantain. Its biological control remains challenging because of the robustness and cryptic behaviour of the adult and the hidden development of larval stages. Researchers therefore tend to favour conservation biological control of this pest. The commonly used methods for measuring the effects of natural enemies on the regulation of this pest focus on invertebrates and may underestimate the role of vertebrates on biological control. Using cameras, we recorded the predation of sentinel adult weevils in banana plots in La Réunion island that differed in weevil infestation levels and in animal biodiversity. To facilitate image analysis, we used background subtraction to isolate moving parts of image sequences and thus detect predators and predation events. Our cameras recorded only vertebrates as predators of adult banana weevils. The most important predator appeared to be the Asian shrew (Suncus murinus), which was responsible for 67% of the predation events. Other predators included the house mouse (Mus musculus), the oriental garden lizard (Calotes versicolor), and the guttural toad (Sclerophrys gutturalis). The exact time of predation events were determined from the images metadata. It was thus possible to identify predator foraging periods that coincided with activity of adult weevils. Our results confirm that images provide useful information for biological and ecological studies. Along with other recent studies, our results suggest that the role of vertebrates in biological control may be underestimated. Based on these results, we advocate for several management implications such as the installation of hedges, grasslands, and ponds to favour these vertebrate predators of the banana weevil, possibly also favouring other vertebrate and invertebrate natural enemies.

Automated image analysis and deep learning tools such as object detection models are being used increasingly by biologists. However, biological datasets often have constraints that are challenging for the use of deep learning. Classes are often imbalanced, similar, or too few for robust learning. In this paper we present a robust method relying on hierarchical classification to perform very small object detection. We illustrate our results on a custom dataset featuring 22 classes of arthropods used to study biodiversity. This dataset shows several constraints that are frequent when using deep learning on biological data with a high class imbalance, some classes learned on only a few training examples and a high similarity between classes. We propose to first perform detection at a super-class level, before performing a detailed classification at a class level. We compare the obtained results with our proposed method to a global detector, trained without hierarchical classification. Our method succeeds in obtaining a mAP of 75 %, while the global detector only achieves a mAP of 48 %. Moreover, our method shows high precision even on classes with the less train examples. Confusions between classes with our method are fewer and are of a lesser impact. We are able achieve a more robust object classification with the use of our proposed method. This method can also enable better control on the model's output which can be particularly valuable when handling ecological, biological or medical data for example.