Scale-Adaptive and Mask Refinement Modules for Accurate Alluvial Fan Boundary Detection in Remote Sensing Data

Hongying Zhou¹Suhong Liu^1*Chengkai Zhou²Zhiguo Ma³

• ¹Beijing Normal University, Beijing, China
• ²Power China Urban Planning and Design Institute Co., Ltd., Guangzhou, China
• ³Research Institute of Petroleum Exploration and Development, Beijing, China

Alluvial fans, as typical depositional alluvial fans in arid regions, play a critical role in geomorphic evolution analysis, hydrological modeling, and land-use planning. Due to their irregular morphological structures and multi-scale characteristics, conventional remote sensing methods face significant limitations in spatial modeling and boundary delineation. To address these challenges, a multi-module enhanced MASK R-CNN framework is proposed in this study, which integrates topographic and spectral information for precise alluvial fan recognition and refined boundary segmentation. The architecture incorporates a Topographic-Spectral Fusion (TSF) input module, a Scale-Adaptive Module (SAM), and a Mask-Boundary Refinement (MBR) module to jointly enhance recognition accuracy and structural detail preservation. Experiments conducted on multi-source remote sensing imagery and terrain data demonstrate that the proposed method achieves an accuracy of $91.7\%$, a precision of $89.8\%$, a recall of $88.5\%$, and an F1-score of $89.1\%$ in full-region classification tasks. For segmentation evaluation, the model attains a mean intersection over union (mIoU) of $81.5\%$ and a boundary F1-score of $80.4\%$. Ablation studies validate the effectiveness of each component, with the TSF module significantly enhancing spatial-structural modeling capability, and the MBR module improving boundary-fitting performance. Moreover, scale-consistency analysis indicates that the model maintains robust performance across fan size categories, with the minimum false negative rate reaching as low as $3.9\%$. These results demonstrate the method's potential as a generalized and transferable approach for complex alluvial fans recognition in arid regions, offering both theoretical value and practical applicability.