Enhancing Building Information Extraction from Remote Sensing Images through Reinforcement Learning from AI Feedback

Deep learning is a standard approach in remote sensing for building information extraction, but it faces challenges like sparse data, large obstacles, and complex light-shadow interactions. To address these, we propose a novel reinforcement learning approach, Reinforcement Learning from AI Feedback (RLAIF), which enhances extraction by using a customized reward network with a deep multi-scale attention mechanism. We also improve the deep deterministic policy gradient algorithm with AI feedback, reducing iterations and boosting robustness. Experimental results show our approach reduces cross-entropy loss per pixel by 3.52% compared to supervised and 7.43% compared to unsupervised methods, while improving recall and accuracy, solidifying its competitive advantage.