Multi-Agent Framework for Inferencing Activities of Objects in Satellite Image

Inferring object activities from high-resolution satellite imagery is a critical task in remote sensing. However, improving activity inference performance typically requires large-scale domain-specific data acquisition and training, which incurs significant time and cost. When the application domain involves national defense and security, data access is further constrained by security classification systems, and inference performance depends on precise annotations by expert analysts rather than simple labeling. Moreover, conventional single-model static inference has structural limitations in capturing complex inter-object interactions and dynamic contexts. This study proposes an experience-based adaptive multi-agent collaboration framework that enhances accuracy and reliability solely through structural optimization at inference time and multi-agent collaboration, without domain data fine-tuning or large-scale retraining. The proposed system comprises five specialized artificial intelligence (AI) agents responsible for perception, vision-language context extraction, hypothesis generation, criticism, and supervision. Through an experience-based adaptive loop that iteratively performs hypothesis generation and critical verification, the framework progressively improves inference accuracy and reliability. Experimental results using the AI-hub satellite image object detection dataset demonstrate that the framework achieves Top-1 accuracy of 67.1%, Top-3 accuracy of 89.5%, Macro-F1 of 0.516, Brier score of 0.516, and expected calibration error (ECE) of 0.018, dominating all five baselines across every metric and reducing the calibration error by approximately one order of magnitude.