High-Precision Remote Sensing Image Change Detection Based on Image Style Unification and Feature Extraction Optimization

Jinmiao Zhao, Zelin Shi, Chuang Yu, Yunpeng Liu

Abstract: Change detection in remote sensing images is crucial for land monitoring, urban planning, and disaster assessment. However, real-world challenges such as style discrepancies between temporal images, sparse change regions, and blurred boundaries hinder detection accuracy. To address these challenges, this paper proposes a high-precision change detection scheme based on image style unification and feature extraction optimization. First, an Image Style Unification (ISU) strategy is proposed to mitigate feature shifts caused by inconsistent imaging styles by aligning low-frequency components, thereby enhancing the model’s robustness in detecting change regions. Second, to improve segmentation accuracy, we replace the conventional encoder with a high-performance backbone and design a hybrid loss function that jointly optimizes classification and boundary precision. Third, an Online Target Region Enhancement (OTRE) strategy is proposed to address the training bias caused by sparse change areas, dynamically enriching training samples and increasing change diversity. Extensive experimental results demonstrate the superior performance of the proposed approach, while ablation studies further confirm the effectiveness and complementarity of each module.

External IDs:dblp:conf/prcv/ZhaoSYL25