Soil Moisture Retrieval And And Drought Monitoring And Vegetation Response Based On Multi-source Remote Sensing And Machine Learning Algorithm

Timely and effective improvement of near-surface soil moisture estimation accuracy and spatial resolution is of great significance to the construction of drought disaster prevention and mitigation capacity in Inner Mongolia grassland and the construction of ecological security barrier in northern Xinjiang.This paper use the multi-source remote sensing data,field measured soil moisture data,meteorological site measured data,soil moisture product data,etc.,from the county wushen banner,Inner Mongolia region and grassland of Inner Mongolia region multiple dimensions,such as optimization and improvement of optical remote sensing soil moisture inversion model,establish a microwave triangle soil moisture index,An optical microwave multi-source remote sensing soil moisture model was constructed using machine learning technology.Second with different downscaling method made an evaluation of the accuracy of soil moisture products,machine learning algorithm is constructed scale transform method of surface soil moisture data,analyzes the soil moisture in the study area and spatial and temporal distribution of soil moisture drought index,based on meteorology,remote sensing,soil relative humidity,and water consumption of grassland plant comprehensive drought index,The responses of hydrothermal conditions and drought to vegetation were revealed.The main research work and achievements of the thesis are:(1)Overall,from R~2 and RMSE,CVDI provides a more accurate spatial change capture than TVDI and OPTRAM.By comparing the correlation between TVDI and OPTRAM and soil moisture content,it was found that R~2 of TVDI and soil moisture content was significantly higher than that of OPTRAM in different months,and RMSE of TVDI and soil moisture content was also significantly lower than that of OPTRAM.This indicates that TVDI is more stable than OPTRAM drought response.(2)Using the measured soil moisture content to evaluate the accuracy of soil moisture index of microwave triangle in different months.The R~2 and RMSE of microwave triangle soil moisture index and soil moisture content in April were 0.77 and 8.62,respectively.The R~2 and RMSE of soil moisture index and soil moisture content in August were 0.62 and 10.04,respectively.This indicates that the microwave triangle soil moisture index has good adaptability to soil water retrieval in the study area.(3)TVDI,σVV,LST,σVH,EVI,SLIA are used as the input variables of four regression tree machine algorithms and two Stacking integration algorithms.Compared with CART,RF and ERT,GBDT algorithm has better learning ability in various groups,thus improving the accuracy of SMC inversion effectively.In April,R~2 of training set increased by 0.06 and RMSE decreased by 2.03%,R~2 of verification set increased by 0.07 and RMSE decreased by2.02%.In August,the maximum R~2 of the training set increased by 0.06 and RMSE decreased by 1.57%,while the maximum R~2 of the verification set increased by 0.05 and RMSE decreased by 1.40%.Stacking1 and Stacking2 models of ensemble learning are compared with the best single model regression tree machine learning model.It was found that R~2 was increased and RMSE was decreased in different months,resulting in a much improved and better Stacking accuracy of the integrated learning model.(4)MODIS optical remote sensing data and SMAP microwave soil moisture data were selected,and VSDI downscaling method,DISPSTCH downscaling method and gradient lifting decision tree machine learning method were used to construct the surface soil moisture data scale transformation method.Data accuracy before and after the scale transformation was compared and analyzed.The soil moisture data of most sites always maintained a good correlation with the 1 km downscale soil moisture products.Meanwhile,we also found similar performance in BIAS,RMSE and ub RMSE.Finally,it was found that the gradient lifting decision tree machine learning method had the highest accuracy,followed by DISPSTCH downscaling method,and VSDI downscaling method.(5)based on the principle of water balance,through the analysis to determine the water requirement of grass and grassland communities in the different periods parameters such as coefficient of evapotranspiration,combined with the standardized precipitation index(SPI),soil(RSM-20 cm)and relative humidity temperature vegetation drought index(TVDI),is proposed to reflect the soil and plant factors and the influence of meteorological conditions comprehensive evaluation index.The annual drought frequency in Inner Mongolia ranged from 16.66%-77.77%,and the drought frequency was different in different regions.The drought frequency in Sunite Left Banner and Erlianhot region was the highest,reaching77.77%,while the drought frequency in Etoki Front Banner was the lowest,only 16.66%.(6)From 1982 to 2020,the average annual decrease of precipitation in Inner Mongolia was 0.357 mm/10a,and the average annual increase of temperature was 0.243°C/10a.From1982 to 2020,the land surface vegetation coverage showed an increasing trend,and the moderate and slight improvement areas accounted for 0.141%and 13.267%of the total area,respectively.The area of degraded area was much smaller than that of improved area,and the area of mildly degraded area accounted for 3.341%of the total area.The overall change of drought is wet-dry-wet,and the impact of land use types on drought is different from different land use types.Grassland is the most sensitive land use type to SPEI in NDVI.In terms of soil texture type,the correlation coefficients between SPEI and NDVI on multiple time scales increased with the increase of sand content,while the correlation coefficients between SPEI and NDVI decreased with the increase of clay content.However,the proportion of loam soil on the correlation coefficient of SPEI and NDVI is closer to clay.The vegetation loss rate under different drought stress was studied by using the probability assessment model of Copula function,and it was found that the NDVI loss rate increased with the increase of drought degree.The effect of drought on NDVI loss probability decreased with the increase of NDVI quantile.