Optimized Reinforcement Learning-Driven Model for Remote Sensing Change Detection

TL;DR

This paper introduces a new remote sensing change detection framework using reinforcement learning to improve accuracy by iteratively refining predictions.

Contribution

A novel feedback-driven CD framework combining U-Net and reinforcement learning for adaptive error correction in change detection.

Findings

The proposed RL refinement increases mIoU by 3.07 to 6.13 points across four datasets.

The framework improves boundary fidelity and suppresses pseudo-changes caused by shadows and illumination variations.

The RL module is adaptable and shows consistent gains when integrated into different CD backbones.

Abstract

In recent years, deep learning has driven remarkable progress in remote sensing change detection (CD); however, practical deployment is still hindered by two limitations. First, CD results are easily degraded by imaging-induced uncertainties—mixed pixels and blurred boundaries, radiometric inconsistencies (e.g., shadows and seasonal illumination changes), and slight residual misregistration—leading to pseudo-changes and fragmented boundaries. Second, prevailing methods follow a static one-pass inference paradigm and lack an explicit feedback mechanism for adaptive error correction, which weakens generalization in complex or unseen scenes. To address these issues, we propose a feedback-driven CD framework that integrates a dual-branch U-Net with deep reinforcement learning (RL) for pixel-level probabilistic iterative refinement of an initial change probability map. The backbone produces a preliminary posterior estimate of change likelihood from multi-scale bi-temporal features, while a PPO-based RL agent formulates refinement as a Markov decision process. The agent leverages a state representation that fuses multi-scale features, prediction confidence/uncertainty, and spatial consistency cues (e.g., neighborhood coherence and edge responses) to apply multi-step corrective actions. From an imaging and interpretation perspective, the RL module can be viewed as a learnable, self-adaptive imaging optimization mechanism: for high-risk regions affected by blurred boundaries, radiometric inconsistencies, and local misalignment, the agent performs feedback-driven multi-step corrections to improve boundary fidelity and spatial coherence while suppressing pseudo-changes caused by shadows and illumination variations. Experiments on four datasets (CDD, SYSU-CD, PVCD, and BRIGHT) verify consistent improvements. Using SiamU-Net as an example, the proposed RL refinement increases mIoU by 3.07, 2.54, 6.13, and 3.1 points on CDD, SYSU-CD, PVCD, and BRIGHT, respectively, with similarly consistent gains observed when the same RL module is integrated into other representative CD backbones.

Figures