Mapping Forage Potential: A Satellite-Based Foodscape Model for Wildlife Management
AI & Data
Semester programme:AI, Machine Learning & Data
Research group:Sustainable Data & AI Application
Project group members:Mohammed Aslan
Priyanka Darbari
Project description
How can satellite imagery data and AI be effectively used to analyze and understand the factors influencing wildlife foodscape patterns, such as vegetation quantity, quality, and land cover?
Context
This project operates at the intersection of environmental science, remote sensing, and artificial intelligence, focusing on wildlife ecology and conservation. Using satellite imagery and machine learning, it maps vegetation types, assesses forage availability, and predicts animal food resources across natural landscapes. The goal is to support evidence-based conservation strategies by translating complex spatial data into actionable insights for park managers, researchers and ecologists.
Results
The main product is a real-time web prototype that maps and predicts wildlife foodscape patterns using Sentinel-2 satellite imagery and gradient boosting models. It serves two Dutch nature reserves (Kempenbroek and Zuid-Kennemerland), each with four model types: land cover classification, vegetation quantity forecast, vegetation quality forecast, and experimental forage estimation. Regression models use a spatial block split for geographic generalization; the classifier uses a scene-based temporal split. The prototype runs as a browser-based application (FastAPI backend, React frontend, interactive Leaflet map). It supports local preprocessed data and live cloud API fetch through backend (Google Earth Engine, Sentinel Hub), and outputs colour-coded map overlays with per-class area statistics in km². Validation consisted of 28 end-to-end test cases covering startup, predictions, date handling, error handling, and UI behaviour, all passed. The system is deployed as a Windows NSSM service.
Validation on a fully unseen year proved more informative than within-year splits for assessing temporal generalization. Park-specific training appears more reliable than generic models due to landscape heterogeneity, and the forage estimation output remains only experimental.
Positioning corresponds to TRL 5, core models and the web prototype are validated in a relevant environment (real parks, real satellite data, real deployment infrastructure), but are not yet fully operational at production scale.