| Poaching and ecosystem degradation continue to be severe threats to the Amazon Rainforest. To support long-term restoration and intervention planning, this work proposes an approach that integrates multi-agent reinforcement learning (MARL) to simulate an ecosystem and graph neural networks (GNN) to detect real-time poaching. The MARL environment simulates complex interactions among key agents of the ecosystem(flora, fauna, soil, climate and water) and gauges the impacts of ecological interventions on the ecosystem health. For poaching detection, objects of interest (campfires, vehicles and waterbodies) are detected and spatial analysis is performed on the satellite images, using YOLOv8 models and GNN. Streaming satellite images, including Sentinel-2 imagery, are processed using Apache Kafka. Results indicate that the proposed system effectively models realistic interactions in the ecosystem, predicts the long-term effects of interventions, and correctly identifies poaching hotspots with additional processing techniques such as image tiling and blue-ratio filtering. The purpose of the proposed system is to help policymakers and conservationists analyze the optimal conservation strategies, while also monitoring poaching in real-time. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.