Abstract: In the black soil region of Northeast China, gully erosion is severe and widespread. Currently, gully monitoring in this area relies predominantly on manual interpretation, highlighting the urgent need for a rapid extraction method. This study selects the Mashezi River Basin in Binxian country, Heilongjiang province, a region heavily affected by gully erosion, as study area. Utilizing GF-7 satellite imagery and comparing with manual interpretation results, the accuracy of three automatic gully extraction methods is evaluated: flow-directional detection, machine learning and deep learning. The findings are as follows: ① The flow-directional detection method depends on high-precision topographic data. The vertical accuracy of topographic data generated from GF-7 stereo images is poor, resulting in an overall extraction accuracy of only 6.7%, and this method is unable to automatically extract permanent gullies and ephemeral gullies from GF-7. ②The machine learning approach requires manual setting of segmentation parameters and design of classification features, limiting its degree of automation. It achieves an overall extraction accuracy of 50.7%, with a precision of 83.1% for permanent gullies and only 9.2% for ephemeral gullies. ③The deep learning method adopts an end-to-end approach, without the need to design feature extractors. It offers a high degree of automation with an overall extraction accuracy of 60.8%, achieving 68.1% accuracy in identifying permanent gullies and 69.7% in recognizing ephemeral gullies.

Key words: black soil region in Northeast China, GF-7 satellite image, flow-directional detection, machine learning, deep learning