Objective This paper optimizes landslide samples based on landslide activity to improve the accuracy of landslide susceptibility evaluation.

Methods The terrain and landforms in the upper of Jinsha River are complex, with strong tectonic activity and the developed landslide disasters. The Baiyu−Batang section of the upper Jinsha River is selected as the key research area, and remote sensing interpretation, InSAR deformation detection, and field investigation techniques are used to identify and analyze landslide activity. All landslides were divided into two datasets: A (active landslides) and B (active landslides and inactive landslides). Eight factors, such as elevation, slope angle, slope direction, engineering geological units, distance to fault, seismic peak ground acceleration, distance to river and NDVI, were selected to complete the landslide susceptibility evaluation by weighted information model.

Results The results show that the AUC based on A and B datasets are 0.855 and 0.810, respectively, indicating that satisfied landslide susceptibility results have been achieved. The very high and high landslide susceptibility is mainly distributed along the Jinsha River and Jiangqu River, and show an obvious band distribution trend along water systems. The middle landslide susceptibility is mainly distributed in the areas between the longitudinal valleys, and the low landslide susceptibility is mainly distributed in flat areas.

Conclusions The accuracy of landslide susceptibility based on A dataset is higher than that of B dataset, and the identification ability of very high and high landslide susceptibility areas is relatively improved. So, landslide activity can effectively improve the landslide susceptibility accuracy, and is an important factor to be considered in the landslide susceptibility evaluation model. The proposed study ideas and methods provide an important reference for promoting landslide susceptibility evaluation in alpine gorge areas.