Deep learning delineates alluvial fans driven by topographic knowledge and imagery

Abstract

An alluvial fan is one of the most typical sedimentary landforms formed by fluvial and depositional geomorphic processes on the Earth’s surface. Extracting the boundaries of alluvial fans is a key procedure for understanding their formation mechanisms and geomorphic processes. In this study, we proposed a method that integrates the topographic characteristics of alluvial fans from a Sentinel-2 imagery and SRTM digital elevation model into an improved deep-learning segmentation model (Mask R-CNN) for alluvial fan extraction. We tested the validity of our method in two representative sample areas in the Great Basin and Mojave Desert regions of the western United States. Results indicated that the method can achieve satisfactory extraction results in these areas and has superior performance over traditional methods, with an F1-score of 91.53% versus 70.92% (mean-shift segmentation) and 70.38% (radial profile). In addition, the relationship between alluvial fans and their corresponding catchments was examined, suggesting that catchment area, slope, relief and rainfall patterns influence sediment transport, deposition and the geomorphological evolution of alluvial fans. Furthermore, the microtopographic features of alluvial fans revealed different degrees of geomorphic development between the study areas. The difference may be primarily attributed to differences in erosion intensity and deposition frequency. Finally, by designing terrain factors that align with specific landform characteristics, the proposed method can be extended to the extraction of other complex landforms.