Research On The Salt Content Model Of Satellite Remote Sensing Monitoring Soil Based On Data Assimilation Algorithm

The problem of soil salinization is one of the important factors restricting the development of agriculture in China.Accurate monitoring of soil salinity is an important way to deal with the problem of soil salinization.Satellite remote sensing can accurately monitor soil salinity in a large area,but there are problems such as discontinuity in time;the solute transport model can simulate the long-term dynamic change process of soil salinity,but there are problems such as narrow test range;data assimilation algorithm combines the advantages of both and reduce the disadvantages.This study took the soil salt content at different depths in the Shahaoqu irrigation area of the Hetao irrigation area as the research object.By calculating the spectral variables and using PLS-VIP to screen the independent variable combinations at different depths in different months;based on the inversion of satellite remote sensing data and different machine learning models,the inversion results of soil salinity at different depths in different months were obtained;at the same time,based on the Hydrus-1D model,driven by ground data and groundwater data,long-term soil salinity simulation results were obtained;finally,two different assimilation algorithms assimilated the inversion results and simulation results,compares and analyzes them,and finally constructs the optimal longterm time-series soil salinity monitoring model in a wide range and at different depths.The main conclusions of the study were as follows:(1)The inversion model of soil salinity at different depths in different months based on the remote sensing of HMS-1 was constructed.By dividing the measured data of soil salinity into a modeling set and a verification set,and using PLS-VIP to filter spectral bands and spectral indices,the optimal combination of independent variables could be effectively obtained and the amount of model calculation could be reduced,saving time for model construction;the soil salinity monitoring models were established based on BPNN,Cubist and ELM models,among which the ELM model had the highest accuracy(the average IOA was above 0.820,and the average ME was below 0.136g/100g),the Cubist model was second,and the BPNN model was the worst.Therefore,this study finally chose the ELM model as the remote sensing inversion model.(2)A simulation model of soil salinity at different depths with long time series based on Hydrus-1D model was constructed.The model was driven by driving data,and the optimal parameters were determined and validated by using the experimental-estimate-correction method to rate the model parameters,in which the results of both the rate set and the validation set had good accuracy and robustness,and could accurately reflect the variation of actual soil salinity;in the simulation results of different months,the accuracy in May was the highest(the average IOA was 0.905,and the average ME was 0.068g/100g),and the accuracy in June was the lowest.However,the overall trend was that the accuracy decreased with increasing time;in the simulation results of different depths,the depth accuracy of 0～20cm was the highest(the average IOA was 0.873,and the average ME was 0.086g/100g),the accuracy of the depth of 40～60cm was the lowest,and the accuracy decreased with increasing soil test depth.(3)Soil salinity data assimilation models based on PF algorithm and UKF algorithm were constructed.The optimal parameter combinations were determined by sensitivity analysis of the two algorithms.In the PF data assimilation model,the optimal parameter combinations were particle number 450,observation noise 0.06 and random resampling;in the UKF data assimilation model,the optimal parameter combinations were observation noise 0.01 and process noise 0.05;the results of remote sensing inversion,model simulation and the two assimilation models were compared and analyzed.The accuracy of UKF assimilation results was the best(average IOA was 0.862,average ME was 0.100g/100 g,and average NMB was0.040),the IOA of ELM model inversion results was lower and ME was higher,the NMB of Hydrus-1D model simulation results was too high,and the overall accuracy of PF assimilation results was lower than that of UKF assimilation results.Therefore,the UKF assimilation model was finally selected as the final long time series and large scale monitoring model for soil salinity in this study,with a view to providing a theoretical basis for the study of regional management of soil salinity problems.