Biodiversity Abundance

Using computer vision techniques for biodiversity monitoring.

Thanks to the newer CV techniques and models we can develope a strategy for monitoring biodiversity. Identifying, counting or geospatially tracking key species can be now automized by the use of this technology. In specific we are developing two models that can help with this tasks: Segementation model for Tree crown counting and Image Classification model for Key species identifiction.

Tree Crown Segmentation (Grounded SAM Model)

To monitor individual trees, the project employs a Grounded SAM model for tree crown segmentation. This approach combines an object detection model (to localize trees) with Meta AI’s Segment Anything Model (SAM) to delineate each tree’s canopy (“crown”) in aerial or drone imagery. SAM is a state-of-the-art segmentation network pre-trained on over 1 billion annotated masks (TreeSeg—A Toolbox for Fully Automated Tree Crown Segmentation Based on High-Resolution Multispectral UAV Data), which allows it to generalize to new objects – in this case, tree crowns – with minimal extra training. Using this model, high-resolution imagery of the reforested sites is processed to identify and outline each tree. The output is a georeferenced map of individual tree crowns, enabling automated tree counting and size measurement. These crown maps support tracking of tree survival, growth, and canopy cover over time for the planted areas. The segmentation accuracy is high for distinct canopies, with modern instance segmentation techniques achieving F1-scores above 90% in identifying trees from high-resolution images under ideal conditions (Tree crown identification and segmentation results by using 4 different... | Download Scientific Diagram). In practice, the model is fine-tuned and validated on sample plots to ensure it accurately detects even young or closely clustered trees in the Senegal project’s landscapes.

Custom Computer Vision for Biodiveristy Monitoring

Beyond tree crown mapping, the project developed ad-hoc computer vision models to detect other forest conditions and validate field activities. These models analyze imagery (from satellites, UAVs, or ground cameras) for specific targets. For example, one application is detecting planting features (such as the presence of prepared planting pits or protective fencing around saplings) to verify that reforestation work has been carried out in each location. Another is assessing tree health and threats: a trained image classifier can flag signs of stress (e.g., browning canopies indicating possible drought or disease) or detect unauthorized land use (like encroachment or fire scars) within the project zones. Modern object detection frameworks (e.g. YOLO) are leveraged here due to their speed and accuracy – YOLOv8-based models can achieve over 93% precision (mAP) in recognizing individual trees and tree clusters in high-resolution imagery (Tree crown identification and segmentation results by using 4 different... | Download Scientific Diagram). The models are tailored to the Senegalese ecosystem; for instance, they can be trained on images of the native tree species and typical vegetation patterns in the reforestation areas. This ad-hoc CV toolkit acts as an automated “eye on the ground,” validating that planted trees remain and identifying any issues between field visits. Alerts generated by these vision models (such as detecting an area with tree loss) prompt targeted field verification, making monitoring more efficient.