Research On Soil Moisture Inversion Based On Deep Belief Network

Soil moisture is an important physical quantity related to land-gas energy exchange,and it is also an important indicator to monitor land condition such as soil degradation and drought.Therefore,it is of great significance to obtain the continuous and high quality soil moisture information for guiding production practice and carrying out scientific research.With the development of satellite remote sensing technology,the retrieval of soil moisture by remote sensing has become the main technology.Satellite remote sensing technology can quickly obtain the relevant information of soil moisture inversion,with low cost,good spatial-temporal resolution and fast time efficiency.It can meet the requirements of monitoring soil moisture in a large range,and is an important data source for obtaining soil moisture information.It is very difficult to obtain high-precision soil moisture data because of the complex relationship between satellite band information and soil moisture as well as the influence of vegetation coverage,temperature and other environmental factors.In view of the above content,Ningxia region is selected as the research area in this thesis.A method of retrieving soil moisture from the image of MERSI-II of FY-3D by using deep belief network(DBN)is proposed.The model is named as SM-DBN(Soil Moisture-Deep Belief Network).The main work and conclusions of this thesis are as follows:(1)In this thesis,FY-3D image data and observation data of ground stations in Ningxia from January 2018 to December 2019 are collected and acquired.The FY-3D data and ground data are preprocessed by using relevant software.Then the required training data set and test data set are used to train and test the model according to the input requirements of the model.(2)Based on the data of the MERSI-II of FY-3D as well as the correlation between the land surface temperature and soil moisture,a hierarchical inversion strategy is proposed by using the FY-3D data to retrieve the land surface temperature and the land surface temperature data combined with the vegetation index data to retrieve the soil moisture.(3)According to the analysis,the SM-DBN model is built.SM-DBN consists of two subnetworks: one is used to retrieve surface temperature and the other is used to retrieve soil moisture.In the land surface temperature sub network,the 1,2,3,4,24 and 25 bands of the temperature related FY-3D MERSI-II image are used as input data,while the land surface temperature obtained by the ground station is used as the expected output value during model training.In the soil moisture sub network,the input data includes the land surface temperature data generated by the trained surface temperature sub network,normalized vegetation index and enhanced vegetation index.The soil moisture data obtained from the ground station is used as the output label of the training model.(4)Linear regression model and BP neural network are selected as the comparison model to compare with SM-DBN model.The same set of data is used to train and test the three models.The experimental results show that the inversion accuracy of SM-DBN model is 91.8%,which is higher than two contrast models.In this thesis,the SM-DBN model improves the accuracy of soil moisture inversion,greatly reduces the dependence on foreign data,reduces the operation cost of a large-scale monitoring business system and improves the stability,safety and monitoring effectiveness of the system.It is of great significance for improving the meteorological service support ability and the application level of domestic satellites.