DETECTION OF ILLEGAL LANDFILLS ON SATELLITE IMAGERY USING A MULTI-AGENT FRAMEWORK

Abstract

This paper presents an extended feature-array model designed for automated detection of illegal landfills in multispectral aerospace imagery. The method relies on a multi-agent architecture in which each detector agent isolates a specific spectral, textural, or contextual indicator, collectively forming a structured representation of the landfill ecosystem. The feature array consists of ten core components that capture infrastructural, environmental, and physicochemical characteristics of landfill sites, ensuring fully interpretable and transparent decision-making without the need for complex stochastic grammars or manually crafted rule sets.

Independent YOLO-based agents are responsible for detecting waste piles, road segments, and nearby industrial facilities. Their outputs are subsequently merged through a probabilistic aggregation mechanism that incorporates spatial consistency, context-aware evidence, and confidence weighting derived from remote sensing data. This integration significantly improves robustness to noise, enables stable operation on medium-resolution imagery, and makes every reasoning step traceable by explicitly revealing the contribution of each feature to the final classification.

The approach was evaluated on an independent test scene near Taromske, Ukraine, and validated through field inspection, confirming the practical reliability of the multi-agent pipeline. Joint aggregation of heterogeneous detections increased confidence in the primary landfill cluster and reduced false positives, while the structured feature array made it possible to reconstruct the full decision path for expert analysis. The proposed system can be further extended to large-scale environmental monitoring, early identification of emerging waste sites, and detection of broader anthropogenic surface disturbances using additional spectral or thermal indicators.