干旱区研究2026，Vol.43Issue(8)：1615-1629,15.DOI:10.13866/j.azr.2026.08.05
基于LSTM和CA-Markov耦合模型的祁连山大野口流域土壤侵蚀特征模拟
Simulation of soil erosion characteristics in the Dayekou Basin of the Qilian Mountains using a coupled LSTM and CA-Markov model
摘要
Abstract
The Qilian Mountains,a vital ecological barrier and"green water tower"in arid Northwest China,play a critical role in the region's ecological security and water resource balance.However,climate change is intensi-fying both freeze-thaw and water erosion in this alpine environment.Existing simulation studies frequently over-look the non-linear temporal dynamics of precipitation and the synergistic effects between land-use change and erosion processes.Furthermore,these models poorly adapt to high-altitude freeze-thaw conditions,thereby limit-ing their utility for guiding precise regional erosion control measures.To accurately predict soil erosion dynamics in a typical watershed under future climate scenarios,this study focuses on the Dayekou Basin within the Qilian Mountains National Park.We propose a novel framework that integrates a long short-term memory(LSTM)neu-ral network,a cellular automata-Markov(CA-Markov)model,and the sediment delivery ratio module of the In-VEST model.First,the LSTM model was trained on 288 months of precipitation data(2000-2024)to forecast rainfall for the period of 2025-2028.Second,using land use data from five years(2019-2023)and five driving factors(elevation,slope,aspect,rainfall,and distance to rivers),the CA-Markov model simulated land use pat-terns from 2024 to 2028,achieving a Kappa coefficient of 0.85.Finally,the predicted rainfall and simulated land use data were input into the InVEST-SDR module to calculate the soil erosion modulus and classify the erosion intensity.The results indicate that from 2024 to 2028,the average erosion modulus of the watershed is projected to fluctuate between 42.322 and 54.978 t·(hm2·a)-1,with intensity primarily ranging from moderate to strong.Al-though total soil erosion shows minor fluctuations,it remains generally stable.Severe erosion zones,located above 3750 m,account for less than 9%of the total watershed area but contribute over 97%of the total soil loss.Notably,the area of these severe erosion zones shows a continuous decreasing trend.This study validates the ef-fectiveness of the integrated model in predicting erosion for complex mountainous watersheds.It innovatively es-tablishes a full-chain methodology of"rainfall prediction-land use simulation-erosion assessment"that addresses the fragmentation of driving factors in traditional models.We also discuss the potential and challenges of this framework for application in high-altitude freeze-thaw zones.The findings provide a scientific basis for soil and water conservation and ecological protection in the Qilian Mountains National Park.关键词
土壤侵蚀/LSTM神经网络/CA-Markov/InVEST模型/大野口流域Key words
soil erosion/LSTM/CA-Markov/InVEST model/Dayekou Basin引用本文复制引用
任小凤,许尔文,韩士哲,赵维俊,康永德..基于LSTM和CA-Markov耦合模型的祁连山大野口流域土壤侵蚀特征模拟[J].干旱区研究,2026,43(8):1615-1629,15.基金项目
国家自然科学基金联合基金(U22A20592) （U22A20592）
甘肃省科技计划项目(25JRRG027) （25JRRG027）
甘肃省生态文明建设重点研发专项(25YFFG004) （25YFFG004）
甘肃省重点研发计划-国合领域(25YFWG001) （25YFWG001）
