Comparative Study On Agricultural Drought Monitoring In The Loess Plateau Based On Three Machine Learning Methods

Agricultural drought hinders the sustainable development of the ecological environment and agricultural system.Accurate and real-time monitoring of agricultural drought is a world problem which that needs to be solved urgently.The traditional monitoring research on agricultural drought is mainly based on meteorological data with a high accuracy,however,it has great limitations in large-scale drought monitoring.Although the development of remote sensing makes up for this shortcoming,the accuracy of remote sensing data cannot meet the needs.It is promising to use advanced methods to construct a large-area,high-precision comprehensive remote sensing drought monitoring index.Machine learning methods can effectively deal with complex nonlinear problems among drought-inducing factors and are widely used in drought monitoring.However,different machine learning methods are suitable for different study areas in drought monitoring.Therefore,this paper integrates multi-source data such as MODIS,TRMM,GLDAS,besides,reanalysis selects 13 drought-related variables,and correlates them with the Standardized Precipitation Evaporation Index(SPEI-3)on a 3-months timescale based on meteorological data.According to the correlation analysis,the variables of the input model were determined by the level of correlation.Three machine learning methods,which including random forest(RF),support vector machine(SVM)and BP neural network,were used to train models in arid,semi-arid and semi-humid regions of the Loess Plateau,respectively.In order to find the machine learning method with the highest accuracy and the lowest error in different regions,this work comparing the the accuracy and error of all models.Then,a Multivariate Comprehensive Drought Index(MCDI)capable of monitoring agricultural drought in the Loess Plateau was obtained by using the RF machine learning method to train models in different regions.Finally,SPEI-3 index,PDSI and VHI drought index based on meteorological data were selected to test the validity of the MCDI drought index.The MCDI drought index was used to analyze the temporal and spatial trends and characteristics of agricultural drought in the growing season of the Loess Plateau from 2003 to 2020.The following main conclusions are drawn from this study:(1)The result of comparing the coefficients of determination and errors of models trained by three machine learning methods in arid,semi-arid and semi-humid areas,shows that the model trained by random forest method has the highest coefficient of determination and the smallest error in the three areas,which is consistent with the support vector machine and BP neural network.Random Forest method is more suitable for agricultural drought monitoring in the Loess Plateau region.At the same time,we found that the accuracy of the three machine learning methods in the sub-humid area is better than that in the arid and semi-arid areas.(2)The MCDI drought index of the entire study area was obtained by using the random forest method to train models in different regions,and the SPEI-3 index based on meteorological data was selected to test the validity of the MCDI drought index.The results show that the above index have a high correlation.The consistency of the variation trend of the stations is great,indicating that the MCDI drought index can be applied to the actual drought monitoring.The cooperation between the drought monitoring capabilities of MCDI,PDSI,and VHI,shows that MCDI has higher precision in agricultural drought monitoring and has more advantages in local area drought monitoring.(3)The spatial distribution of drought in the agricultural growing season on the Loess Plateau has obvious regional characteristics.From 2003 to 2020,the drought in the study area at the junction of Shanxi and Henan and its surrounding areas continued to increase with time,and the drought in Ningxia and Gansu was alleviated.From 2003 to 2020,the averaged value of MCDI in the agricultural growing season continued to increase with the monthly scale change.The average value of MCDI in April showed a downward trend after 2014.The overall value of MCDI in May was low.From June to August,it increased significantly after 2012,and from September to October,except for individual years,it was higher.(4)The frequency of drought in the agricultural growing season on the Loess Plateau showed an overall trend of increasing at first stage,decreasing at second stage,and then increasing from 2003 to 2020.The frequency of drought was obviously regional.The areas with higher frequency of drought were in Ningxia and Inner Mongolia,which shows a decreasing trend from northwest to southeast.The highest proportion of severe drought and extreme drought were existed in April,the proportion of mild drought and moderate drought was high in May-July and September-October,and the proportion of mild drought and severe drought was higher in August.