Soil Moisture Retrieval Based On Machine Learning

Accurate and fast access to large areas of soil moisture information is very important for agricultural irrigation water management and drought prevention.The traditional soil moisture monitoring method has the disadvantages of high sampling cost and poor spatial representation,for this reason,people are looking at the emerging remote sensing method to monitor soil moisture.Vegetation index method and microwave forward model method are two commonly used methods for monitoring soil moisture by remote sensing,but the vegetation index method model is too simple,the model only considers the effect of vegetation on soil moisture,and the microwave forward model is very complicated and needs to consider many factors,also the SAR radar’s ability to detect vegetation is limited.Therefore,combining two kinds of remote sensing data to jointly retrieval soil moisture has become a new research direction,and how to better combine two kinds of remote sensing data to retrieval soil moisture is the focus of this study.Considering the intricate quantitative relationship between soil backscatter coefficient,vegetation index and soil moisture,and machine learning algorithms can simulate this quantitative relationship well,in the experiment,four machine learning algorithms including linear regression algorithm,support vector machine regression algorithm,AdaBoost regression algorithm and KNN regression algorithm were combined with Landsat 8 image and Sentinel-1 image.The Naqu region was used as the research area to establish an inversion model,evaluate the model effects of the four models and select one of the models to complete the soil moisture retrieval of the entire study area in late July 2015.The research results show that the coefficient of determination of the multiple linear regression model on the training set is 0.621,the root mean square error is 6.26%,and the absolute error is 4.99%.The coefficient of determination of the multiple linear regression model on the test set is 0.612 and the root mean square error is 7.48%,the absolute error is 6.21%;the determination coefficient of the support vector machine regression model on the training set is 0.622,the root mean square error is 6.24%,the absolute error is 5.12%,the determination coefficient R2 of support vector machine regression model on the test set is 0.622,The root mean square error is 7.38%,the absolute error is 6.20%;the coefficient of determination of the AdaBoost regression model on the training set is 0.814,the root mean square error is 4.38%,the absolute error is 3.75%,the coefficient of determination of the AdaBoost regression model on the test set is 0.625,the root mean square error is 7.35%,and the absolute error is 6.13%;the coefficient of determination of the K-nearest neighbor regression model on the training set is 0.624,the root mean square error is 6.23%,and the absolute error is 4.83%.The determination coefficient of the K-nearest neighbor regression model on the test set is 0.633,the root mean square error is 7.27%,and the absolute error is 6.08%.Synthesizing the performance of the model on the training set and the test set,it is concluded that the AdaBoost regression model is most suitable for soil moisture retrieval in the Naqu area.The model has ability to predict soil moisture and performs well on the test set.Subsequently,we used the AdaBoost model combined with Landsat 8 image data and Sentinel-1 image data in late July 2015 to predict the soil moisture distribution in Naqu in late July 2015,and generated a soil moisture distribution map.The analysis results showed that the overall soil moisture in the study area showed a decreasing trend from southeast to northwest.The average soil mositure in the study area was 20.05%,and 23.56%of the area had drought and water shortage.The results are consistent with the climatic and topographic features of the Naqu region,and better reflect the spatial trend of soil moisture distribution in the Naqu region.