Research On Remote Sensing Drought Monitoring Based On Machine Learning

Drought is a water shortage phenomenon caused by long-term water supply and demand imbalance.With the continuous increase of global temperature,the frequency and intensity of drought on global and regional scales are increasing.The study of drought problem has also become one of the hotspots of global research.Drought has the characteristics of long duration,high frequency,wide influence range and serious economic loss,and it tends to cause serious harm to land resources,water resources and agricultural resources.Therefore,accurate detection of drought occurrence is of great significance to the protection of ecological environment and the development of social economy.Due to the complexity of drought and the diversity of influencing factors,the accurate monitoring of drought still faces many problems.In particular,the drought in Inner Mongolia is becoming more and more frequent and aggravated,and the formation and disaster-causing process has certain particularity.The traditional drought monitoring methods can no longer meet the requirements of regional drought monitoring,so more scientific monitoring methods and means are needed.At present,the commonly used data for drought monitoring are weather station monitoring data and remote sensing image data.However,neither of these two data can achieve the requirements of high precision,high coverage and high timeliness at the same time.Therefore,this paper will use the coupling of model meteorological data and remote sensing indicators to construct a remote sensing drought monitoring index,and downscale the drought monitoring indicators through machine learning model training,so as to achieve high-precision and large-scale regional remote sensing drought monitoring.In this paper,the east Ujimqin grassland in Xilingol League of Inner Mongolia Autonomous Region was taken as the research area.The MODIS data,FLDAS data and CHIRPS precipitation model data of each growing season(May to September)from 2001 to 2021 are selected.The remote sensing indexes related to drought were selected from the perspectives of meteorology,vegetation and soil.Four machine learning models were used for training,and the remote sensing drought monitoring index(RDMI)was constructed by pixel-by-pixel processing.The MODIS data with high spatial resolution was used to realize the downscaling of drought monitoring index.Finally,the RDMI index was used to analyze the spatial and temporal variation characteristics of drought trend in the study area.The main conclusions are as follows:(1)In this study,remote sensing data and model data were used to select multiple drought indicators from the three directions of vegetation,soil and atmosphere,and the comprehensive meteorological drought index CI was used for correlation analysis to screen indicators suitable for drought monitoring in the study area.After screening,NDVI,LST,ET,SM,PRE and CI had a high correlation,and the above indicators were selected as independent variables for constructing remote sensing drought monitoring index.(2)The selected drought indexes were normalized as vegetation condition index(VCI),temperature condition index(TCI),soil moisture condition index(SMCI),evapotranspiration condition index(ETCI)and precipitation condition index(PCI).Taking the condition index as the independent variable and CI as the dependent variable,four machine learning models of BP neural network,support vector regression,random forest and convolutional neural network were selected to train the remote sensing drought monitoring index month by month.Through the evaluation of model accuracy and drought monitoring ability,it is found that the random forest model has the best simulation accuracy,and the refined remote sensing drought monitoring index(RDMI)is constructed by using the random forest model.(3)RDMI was used to analyze the drought space-time in the study area.The analysis found that the climate in the study area gradually changed from dry to wet between 2001 and 2021.From 2001 to 2011,the climate in the study area was mild drought,and severe drought occurred in 2007 and 2010.From 2012 to 2021,the climate was relatively humid and there was no drought.Although there was a certain degree of fluctuation in 2017,the overall trend was still wet.With the change of seasons,the trend of climate change from drought to wet in the late summer and early autumn(August and September)was more obvious than that in the late spring and early summer(May-July),which proves that the trend has significant seasonal differences.(4)The RDMI values of meadow steppe and some typical steppe in the eastern part of the study area showed a significant upward trend from May to September.The typical grassland area in the central and western part of the study area showed a significant upward trend only in August and September.This phenomenon proves that there are regional and temporal differences in climate change in different regions.At the same time,the typical steppe area in the west of the study area is more prone to drought than the meadow steppe area in the northeast,indicating that the occurrence of drought is different among different vegetation types.