Research On Quantitative Inversion Of Soil Moisture Based On Multi-source Remote Sensing Dat

Soil moisture is an indispensable component of the earth’s ecosystem,an important factor for agricultural development,and a major environmental factor for crop growth.Therefore,realtime and accurate monitoring of soil moisture over a large area is important for regional agricultural development and drought prediction.Inversion of soil moisture content by coupling active microwave and optical remote sensing is a hot research topic nowadays.However,most of the existing researches are based on empirical and physical models,and the inversion of soil moisture content is performed by a limited number of parameters,and some parameters need to be obtained through a large number of actual measurements,which is time-consuming and cannot meet its timeliness,and also cannot measure the weight of each feature in the model construction as a whole.With the development of machine learning and neural network algorithms,it provides a way to fit the complex nonlinear mapping relationship between feature parameters and soil water content.In this study,a total of 19 feature parameters were extracted from radar and optical images based on Sentinel-1 and Landsat8 image data,and the order of input parameters was determined by normalized summation to construct multiple soil water content inversion models with multiple parameter combinations,and the optimal model was selected.The optimal combination of parameters was selected.The main conclusions of the study are as follows:(1)By normalizing and summing the importance weights of the features derived from the Mean Decrease Accuracy(MDA)feature selection algorithm and the Pearson feature correlation results,the importance and correlation of the variables are taken into account,which is more explanatory for the contribution of the features in the construction of the soil water content inversion model and more suitable for this study.This is the main basis for the input order of model parameters in this study.(2)A cross-sectional comparison of the models shows that the Sparrow Search Algorithm Back Propagation(SSABP)algorithm has the best overall effect in constructing the soil water content inversion model.Secondly,Particle Swarm Optimization Back Propagation(PSOBP)has better accuracy in constructing the inversion model,but it relies on increasing the input features to improve the accuracy of the model at the expense of efficiency;Random Forest(RF)constructs a more stable inversion model for each parameter.Random Forest(RF)inversion model is more stable and less affected by the dimensionality of the parameters,but the model accuracy is average;Back Propagation(BP)model is extremely unstable and the overall accuracy of the model is poor.(3)It was found that the SSABP6 model constructed by using the first six parameters,Swir2,EVI,VH*VV,Nir,Blue and Swir1 as input features,was the best soil water content inversion model for this study,with R~2 of 0.9272 and RMSE of 0.0076 for the modeling group and R~2 of 0.8074 and RMSE of 0.0115 for the testing group.Compared with other models,the inversion performance is better and the inversion results are more consistent with the actual spatial distribution.