Abstract:

Erosion caused by soil water is widespread in the hillgully Loess Plateau of central Gansu,China,but it is difficult to observe and predict.The objectives of this study were to establish a predictive model for soil water erosion based on a partial least squares algorithm (PLSR) and long short-term memory (LSTM),and to evaluate the model’s effectiveness using a back-propagation neural network,a recurrent neural network,and a LSTM neural network.The results were verified with experimental data from a runoff field of artificial grassland in the soil and water conservation experimental stations in Anjiagou,Dingxi City,from 2005 to 2016 (1-12 months).Input factors for the model were monthly precipitation,monthly erosive precipitation,monthly runoff,monthly rainfall intensity,runoff area,slope of runoff field,soil sand content,and clay sand content ,from which the model predicted monthly soil water erosion.The results revealed that the original eight input factors could be reduced to four by the PLSR algorithm in the prediction model,alleviating the requirement of the LSTM neural network model for a large number of samples.Our method improved both the prediction precision and the convergence rate of the soil water erosion model in artificial grassland by combining a PLSR algorithm with a LSTM neural network.The average relative error of our prediction results was less than 4%,and the correlation coefficients were higher for our model than for the three neural network models used for comparison.In addition,iteration times,root mean square error,and mean absolute error were lower than for the other neural network models.We found that slope had a significant effect on soil water erosion,but only when the monthly rainfall was more than 25 mm.Our PLSR-LSTM neural network model could be used to accurately predict soil water erosion in artificial grassland in the hill-gully Loess Plateau of central Gansu,and it could provide a new path toward accurate prediction of soil erosion in this area.

Key words: hill-gully loess plateau; artificial grassland, long short-term memory (LSTM) recurrent neural network, soil water erosion prediction