A Continuous Perspective-Based Strategy for Deep Learning in Semantic Segmentation for Remote Sensing Fault Mapping

Faults are critical geological features visible in remote sensing imagery, playing essential roles in earthquake studies, resource exploration, and disaster prevention. Traditional automatic extraction methods for these linear features suffer from low accuracy, discontinuous results, and a reliance on manual thresholding. This study proposes a strategy based on continuous perspectives to enhance deep learning for semantic segmentation of faults in optical remote sensing images. The model simulates expert multi-view observation by inputting a target image patch alongside its eight surrounding patches, incorporating a center-masked loss strategy and multi-scale attention modules in a U-Net-like architecture to capture directional and positional features across scales. Subsequently, a post-processing module utilizes an ensemble inference strategy derived from five-fold training to select possible fault areas (maximum probability ≥ 0.5) based on high-probability key points (mean probability ≥ 0.5), followed by Zhang-Suen thinning to generate continuous linear outputs, eliminating manual thresholds. Experiments on Landsat 8 OLI data from Inner Mongolia, China, demonstrate that the proposed model achieves a 25.56% increase in recall compared to prevalent semantic segmentation networks. The method successfully detected 74.04% of the labeled quasi-linear fault segments exceeding 800 m. This performance is attributed to the continuous perspective-based strategy and the post-processing module, which collectively improve feature linearity and fault localization accuracy. Ultimately, the proposed method serves as a valuable reference for the intelligent interpretation of other complex linear geological features in remote sensing applications.