Automatic extraction of mountain river information from multiple Chinese high-resolution remote sensing satellite images

High-resolution geomorphic information from remote sensing images is a key part of mountain river research. However, the complete information about narrow rivers is difficult to extract automatically and accurately from complex backgrounds, especially with mountain shadows. This research uses a random forest (RF) algorithm with an artificial neural network (ANN), RF-ANN, to analyze remote sensing images. This method supports parallel operations and reduces the scale of the infrared data for noise removal to achieve pixel-level extraction of the river surfaces. The RivWidthCloud (RWC) method is improved using Laplacian and edge algorithms for automatic extraction of the bankfull river widths. The improved RWC method is generalizable since it does not require setting the discriminant threshold manually. The method is then applied to the Huangfuchuan River Basin on the Loess Plateau, China using images from the Chinese GF-1 and ZY-3 satellites as the primary data source to extract the river surfaces and widths of the rivers above level 2. The results show that the RF-ANN method has a 94.7% accuracy for extracting river surfaces. The bankfull river width extraction error is 1.07 m (about 0.5 pixels) and the minimum river width extracted by these methods is 6.1 m (about 3 pixels). R² is 0.93 and the root men square errors (RMSE) is 1.52 for fitting the extracted river widths and the test river widths. For small rivers narrower than 10 m, the extraction error is 18.5%, for widths from 10 to 30 m the error is 8.8%, for widths from 30 to 90 m the error is 2.0%, and for rivers wider than 90 m the error is 0.7%. These results provide accurate datasets for watershed topography research in mountainous and other complex topographic regions.