DERL: A Differential Evolution-Reinforcement Learning Framework for Efficient Hyperparameter Optimization in Hyperspectral Image Classification

The performance of hyperspectral image classifiers is highly sensitive to the configuration of their hyperparameters. However, manual tuning is inefficient and impractical, especially for high-dimensional and hybrid hyperparameter spaces. Therefore, Hyperparameter optimization (HPO) has emerged as a critical technique for improving model performance. However, existing HPO methods frequently suffer from limited efficiency and suboptimal performance when navigating such complex spaces. To address these challenges, this paper introduces a novel HPO framework by integrating Differential Evolution with Reinforcement Learning (DERL). In this framework, a Deep Qlearning-based RL agent is embedded into the DE optimization process to dynamically adjust key DE parameters and mutation strategies in real-time, based on the evolving optimization state. This bidirectional coupling enables the algorithm to intelligently steer the search trajectory, balancing exploration and exploitation more effectively. Extensive experiments conducted on four hyperspectral datasets demonstrate that DERL consistently outperforms both advanced and mainstream HPO methods on relevant metrics. Ablation studies validate the contributions of individual components, highlighting the importance of adaptive mutation strategies and two-way information transfer between DE and RL. Furthermore, generalization experiments across different classifiers affirm the robustness and versatility of the proposed method. This study presents an efficient, and adaptive HPO strategy tailored for complex hyperspectral image classification tasks, and underscores the effectiveness of integrating evolutionary algorithms with reinforcement learning to enable automated and robust model optimization.