Francesca Drăguț
With an MSc in Geo-information Science and a BSc in Computer Science, I dedicate my skills to ecology and wildlife conservation efforts, using Remote Sensing, AI, and GIS.
Sessions
This presentation summarizes a thesis study which developed and compared Random Forest models for fish habitat mapping. The case study focused on a South Indian wildlife sanctuary inhabited by Humpback Mahseers, a critically endangered fish species.
The workflow relied on open-source R packages, and preprocessing included Sentinel-1 data for riverbed delineation. Models were trained on PlanetScope (PS) spectral and textural features, using training samples derived from UAV imagery and local expert knowledge. Three Random Forest models were trained on one river stretch and tested on two others, using: (1) all PS bands and textures, (2) only textures on all PS bands, (3) red and near-infrared bands and textures. Validation data was derived from Google Earth Pro Airbus imagery.
The model trained only on textures achieved the best performance, with 0.56 recall and precision, and 0.89 overall accuracy. This research shows a replicable and low-cost mapping workflow which provides local experts with accessible tools to interpret habitat maps and relate them to how endangered species use them. While PS data was available in this study for research and educational use, integrating additional open-source data - or identifying suitable open alternatives - could enhance conservation efforts in data-scarce and resource-limited environments.