The prediction of cropland gully length density in the Songnen black soil region of Northeast China

We built our model using Random Forest algorithm based on high quality gully density samples in 853 watersheds and 37 regional distributed factors. The samples were obtained through visual interpretation of sub-meter imagery and validated in 55 watersheds using centimeter-resolution UAV imagery and field surveys. Then the model was applied to 85,516 sub-watersheds covering the whole domain. Our model demonstrated good predictive accuracy at regional scale, with a Nash-Sutcliffe Efficiency (NSE) of 0.6 for gully density prediction.