GF-1 Satellite Remote Sensing Monitoring Method For Soil Salinization Based On Data Assimilation

Soil salinization is a serious restrictive factor affecting sustainable agricultural development.Satellite remote sensing technology can monitor soil salinization on a large scale,but most of the previous studies have not systematically consider the impact of vegetation coverage on monitoring accuracy.The data assimilation method based on satellite remote sensing combines the respective advantages of observation and simulation to obtain more accurate prediction result,but this method is rarely used in salinization motoring.Therefore,we carried out soil collection experiment in time series in Jiefangzha Irrigation Area of Hetao Irrigation District,Inner Mongolia.Firstly,we used GF-1 satellite imagery and extracted different types of spectral covariates,which were taken as independent variable,used machine learning algorithm to inverse soil salt content(SSC)and obtained the soil salinity observations under different fractional vegetation coverage(FVC).Then,we used Hydrus-1D software to simulate the movement of soil salt.According to the actual situation of the irrigation area and previous experience,we determined various parameters combining with indoor and outdoor experiments,and output the simulated salt content of the soil profile.Finally,we used the particle filter algorithm to fuse soil salinity observations with the simulated salt content and obtain the data assimilation(DA)value.The main results obtained are as follows:(1)The SSC inversion model was constructed by GF-1 satellite remote sensing under different vegetation coverage.Firstly,we filtered out 19 spectral covariates with best subset selection method,and found that the contribution of vegetation spectral index to vegetation dataset is greater and it of salt index to the bare land dataset is greater.Then,three algorithms,Partial Least Squares(PLSR),Cubist,and Extreme Learning Machine(ELM),were used to construct SSC inversion models with each FVC at three different depths.Comparing the model performance we can find that the effect of Cubist model is best,followed by ELM,and PLSR the worst;the model accuracy with vegetation division is significantly higher than that without vegetation division.Finally,we used the best inversion model to draw soil salinity distribution maps,and we could find that dividing vegetation coverage significantly improved the accuracy of the distribution map,making the interpretation of soil salinization by GF-1 satellite more accurate in the test area.(2)The SSC simulation model during the whole crop period was constructed by water-salt transport model of Hydrus-1D.The obtained soil salt simulation results at surface showed obvious decline nodes around 75 d and 100 d,which were mainly affected by factors such as irrigation and precipitation;the SSC simulation result at deep soil was decreased,and it was mainly due to strong crop transpiration and evaporation,which caused the upward movement of salt with water;the SSC simulation result at the depth of20-40 cm was affected by a variety of factors,resulting in a continuous and gentle upward trend of salinity.We evaluated the simulation results with three indicators,and found that the error between estimated soil salt content and measured soil salt content is relatively small,and the predict performance of the depth at 20-40 cm is the better one,followed by0-20 cm,the performance at 40-60 cm is worse.(3)The SSC assimilation model with particle filter algorithm was constructed by fusing the satellite remote sensing and Hydrus-1D simulation.The results of assimilation were closer to the measured values in terms of the trend and degree of soil salinization.The error of data assimilation process is 29%?57% lower than that of Hydrus-1D simulation process at three different depths.Among the three depths,the best assimilation effect is20-40 cm,with RMSE of 0.0422,followed by 40-60 cm,and the worst is 0-20 cm,with RMSE of 0.0713.The sensitivity of particle number and observation error to the assimilation process was further explored,and the analysis shows that among the three depths,the best assimilation effect is 20-40 cm,with RMSE of 0.0422,followed by40-60 cm,and the worst is 0-20 cm,with RMSE of 0.0713.In conclusion,using data assimilation method can significantly improve the monitoring accuracy of soil salinization.