Smart Drones, Smarter Learning: Federated Self-learning Minimal Learning Machine Classifier for Real-Time Hyperspectral Image Classification

Abstract. This paper presents a framework for real-time hyperspectral image classification using federated self-learning Minimal Learning Machines (SL-MLM) and trajectory-optimized UAV swarms. The proposed method enables on-board model training and prediction with low computational cost, supporting asynchronous collaboration between UAVs via adaptive Kalman filter-based model fusion. To optimize scanning efficiency, we integrate deep reinforcement learning-based trajectory planning using a Multi-Agent Deep Q-Network (MADQN), minimizing total flight duration and improving energy efficiency. Experimental results on the Salinas-A hyperspectral dataset demonstrate that our federated SL-MLM achieves high classification accuracy with minimal labeled data and communication overhead. The approach supports scalable and distributed remote sensing applications in bandwidth- and resource-constrained UAV environments.