Optimized Reinforcement Learning-Driven Model for Remote Sensing Change Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning has advanced remote sensing change detection (CD), but practical use is limited by imaging uncertainties like mixed pixels, blurred boundaries, and radiometric inconsistencies.
• Current CD methods lack adaptive error correction, hindering generalization in complex or unseen environments.

Purpose of the Study:

• To develop a feedback-driven CD framework that addresses limitations in current methods by integrating deep reinforcement learning (RL) for iterative refinement.
• To improve the accuracy and generalization of CD by adaptively correcting errors caused by imaging uncertainties.

Main Methods:

• A dual-branch U-Net architecture is combined with a Proximal Policy Optimization (PPO)-based RL agent for pixel-level probabilistic iterative refinement of change probability maps.
• The RL agent uses a state representation fusing multi-scale features, prediction confidence/uncertainty, and spatial consistency cues to perform multi-step corrective actions.

Main Results:

• The proposed RL refinement module consistently improved performance across four datasets (CDD, SYSU-CD, PVCD, BRIGHT).
• For example, integrating the RL module with SiamU-Net increased mean Intersection over Union (mIoU) by 3.07, 2.54, 6.13, and 3.1 points on the respective datasets.
• Similar gains were observed when the RL module was applied to other CD backbones, demonstrating its versatility.

Conclusions:

• The feedback-driven CD framework effectively enhances boundary fidelity and spatial coherence while suppressing pseudo-changes caused by imaging artifacts.
• The RL module acts as a learnable, self-adaptive mechanism for optimizing imaging interpretation in remote sensing applications.