Weakly Supervised Leaf Disease Segmentation via Reinforcement Learning-Based Prompt Optimization for SAM

Abstract

Accurate segmentation of leaf lesions is essential for crop disease monitoring and precision agriculture. However, most existing methods rely on pixel-level annotations, which are costly and difficult to obtain in real agricultural scenarios. Weakly supervised approaches alleviate this requirement but often suffer from incomplete lesion localization and background interference. In addition, recent foundation segmentation models demonstrate strong generalization ability, yet their performance heavily depends on prompt quality, which is difficult to design automatically. To address these challenges, this paper proposes a weakly supervised lesion segmentation framework that integrates probabilistic feature disentanglement and reinforcement learning-based prompt optimization. First, a probabilistic class activation mapping (CAM) is introduced to model lesion confidence and guide feature disentanglement, effectively suppressing background noise. Second, a dual-loss constraint is designed to improve pseudo-mask quality by enforcing both region consistency and structural regularity. Finally, a prompt optimization strategy is developed to automatically select informative prompts, enabling a foundation segmentation model to refine lesion boundaries without human interaction. Experimental results on multiple agricultural datasets demonstrate that the proposed method achieves superior performance compared with existing weakly supervised baselines. Specifically, the method achieves an mIoU of 45.6% and Dice score of 57.1% on apple leaf disease datasets. The results indicate that the proposed framework can effectively reduce annotation cost while providing reliable segmentation performance, indicating potential for practical agricultural applications.