Research On Soil Moisture Retrieval Based On Multi-source Data

Soil moisture is an important element of the terrestrial surface system.It is not only an important physical quantity for land-atmosphere energy exchange,but also an important indicator for monitoring soil drought.Therefore,it is of great significance to obtain soil moisture information with a wide range and fine spatio-temporal resolution.Satellite remote sensing technology can carry out real-time dynamic monitoring of the ground surface,and can quickly obtain surface information in a wide range,which makes the use of remote sensing inversion as the main method to obtain soil moisture.However,it is still difficult to obtain highprecision soil moisture data through remote sensing technology due to the influence of weather and cloud cover and the existence of environmental factors such as vegetation cover.In view of the above content,this article selects Shandong and Ningxia as the research area.Based on GF-3 microwave remote sensing image,Landsat8 optical remote sensing image and ground observation data,utilizing the advantage of the deep belief network to mine the deep correlation of data,replacing the restricted boltzmann machine with a fully connected layer,a model for inverting soil moisture was constructed,named Soil Moisture Net(SMNET).The main work and conclusions of this paper are as follows:(1)This article obtained the GF-3 microwave remote sensing image,Landsat8 optical remote sensing image and ground observation data from January 2018 to September 2020 in Shandong and Ningxia,and preprocessed with related software.Then according to the input requirements of the model,the required training data set and test data set are made to train and test the model of this article.(2)Based on the deep belief network,the SMNET model is built with the fully connected layer instead of the restricted boltzmann machine as the basic unit.The SMNET model includes four parts: input,feature miner,numerical simulator and output.The feature miner is composed of multiple feature extraction layers and is used to mine the nonlinear relationship between input data.The numerical simulator fits the nonlinear relationship between the feature information extracted by the feature miner to the soil moisture.Among them,the backscatter coefficient and polarization ratio which extracted from the GF-3 image,the enhanced vegetation index and the normalized vegetation index which extracted from the Landsat8 image are used as the input values of the model.The observed soil moisture value obtained from the ground observation site is used as the output value during model training.(3)This paper choose Deep Belief Network,Linear Regression model and Random Forest model as the comparison model.The same data set is used to train and test the SMNET model,Deep Belief Network,Linear Regression model and Random Forest model.The experimental results show that the inversion accuracy of the SMNET model reaches 92.3%,which is higher than the inversion accuracy of the three comparison models.Based on the GF-3 microwave remote sensing image,Landsat8 optical remote sensing image and ground observation data,this paper builds an SMNET model to study soil moisture retrieval,which not only improves the accuracy of soil moisture retrieval,but also satisfies the demand for obtaining soil moisture data with high spatial and temporal resolution and numerical accuracy.The use of domestically produced GF-3 satellite data also reduces the dependence on foreign data and reduces the cost of large-scale monitoring of soil moisture,which is of great significance for improving the level of domestic satellite applications.