Background Jianshui county of Yunnan province is a typical karst faulted basin landform, which is an important type area for the comprehensive control of karst rocky desertification in China. Because of the severe seasonal drought and rocky desertification, the barren water holding capacity of soil is poor and heterogeneity is high, and the prediction of soil water dynamic is difficult. It is of great significance for soil and water conservation and ecological restoration to reveal the process of soil water dynamic change and its influencing factors in this area.

Methods Based on the daily meteorological data from April 16, 2006 to December 1, 2020 and soil moisture data from two different degrees of rocky desertification areas in Jianshui Karst rift basin of Yunnan province, a dynamic prediction model of soil moisture volume based on BP neural network was established for 0-10, 10-20 and 20-30 cm soil layers. Genetic algorithm was used to optimize the weights and thresholds of the model. The default factor method was used for sensitivity analysis to identify the main meteorological factors affecting the prediction of soil water dynamics in this area.

Results The BP neural network model optimized by genetic algorithm was used to predict the soil volume moisture content of mild rocky desertification area and moderate rocky desertification area from September 13, 2019 to December1, 2020. The results showed that the predicted value of the model was close to the measured value. Y_MARE increased by 45% and 63%, Y_RMSE by 3% and 12%, R² by 27% and 17%, respectively. The simulation accuracy of soil water in 20-30 cm depth was improved most obviously, The sensitivity analysis showed that the sensitivity index of soil moisture to rainfall was the highest (1.317-1.735), followed by the sensitivity index to average temperature (0.880 9-1.071 2), followed by atmospheric pressure and solar radiation.

Conclusions The simulation result of genetic algorithm optimization was improved obviously. The results show that the BP neural network model optimized by genetic algorithm can be well applied to soil moisture simulation in rocky desertification area, and the simulation accuracy is significantly improved compared with the non-optimized model. It is proved that the prediction accuracy of soil water in moderate rocky desertification plot is higher than that in mild rocky desertification plot, and the prediction accuracy decreases with the increase of soil depth. Sensitivity analysis was used to determine that precipitation was the main meteorological factor, followed by air temperature. The sensitivity analysis showed that the soil at 0-10 cm surface layer was the most sensitive to meteorological factors, while the soil at > 10-20 cm middle layer was the most sensitive to meteorological factors.