Retrieval Method For Key Parameters Of Water Conservation From Remote Sensing Data

Soil moisture and surface temperature are important physical quantities that characterize energy conversion between various layers of the ecosystem(soil layer,water layer,vegetation layer,atmosphere),and are key parameters reflecting the water conservation status of terrestrial ecosystems.Therefore,soil moisture and temporal and spatial variation of surface temperature is of great significance for assessing the water conservation status of the area.Compared with traditional high-cost ground measurement methods,microwave remote sensing technology can quickly and extensively acquire long-term sequence surface parameters.In the past,researchers have been working on the principle of radiative transmission of passive microwaves to construct surface parameter retrieval algorithms,and combined with empirical statistical models and machine learning methods to improve the retrieval algorithms for different regions and different surface types.With the development of deep learning technology,convolutional neural networks with multi-layer nonlinear transformations for high-complexity data modeling are widely used in various fields,which provides a new opportunity for further improvement of traditional remote sensing parameter retrieval algorithms..Based on the analysis of previous retrieval algorithms,this study summarizes the advantages and limitations of different algorithms,and proposes a Convolutional Neural Network(CNN)algorithm for retrieval of soil moisture and surface temperature to overcome traditional retrieval methods.The defects,improve the accuracy of retrieval,and analyze the temporal and spatial variation characteristics of key parameters of water conservation in China.The main research results are as follows:(1)The accuracy and practicability of CNN soil moisture retrieval were analyzed.The simulation data for the classic AIEM model of the bare land surface,the simulation data of the matrix double(M-D)model for the low vegetation surface,and the soil moisture data of the CLDAS were selected to establish the sample database to retrieve soil moisture.Accuracy analysis under test samples shows that the accuracy of CNN retrieval is lower with the reduction of high frequency channels.This shows that the more channels,the more sufficient the CNN can extract the characteristic information of the microwave radiation signal,and the higher the accuracy of retrieving the soil moisture.In order to make CNN soil moisture retrieval more in line with the actual situation of the surface,this study comprehensively uses model simulation data,CLDAS soil moisture products and reliable ground measurement data samples to form a multi-source training database.The result shows that the root mean square error(RMSE)between the CNN retrieval results and the ground observation data based on the multi-source database framework is 0.0384 m~3/m~3 with high precision(R~2=0.8945).(2)The applicability and accuracy of CNN surface temperature inversion algorithm in different combinations and different regions were analyzed.The research shows that the combination of 12 V/H channels of passive microwave makes the CNN retrieval model the most stable and accurate.CNN retrieval algorithms based on different regions show that CNN has higher accuracy in large bare-surface areas.The fitted line of the CNN retrieval data and the ground measured data of the test sample is generally close to 1:1,indicating that the land surface temperature of the CNN retrieval can maintain the overall spatial variation with the actual value,indicating that the CNN can be applied to the land surface temperature retrieval.The accuracy verification results that the correlation coefficient R~2 is 0.987,the RMSE is 2.69K,and the average relative error is 2.57K,which indicates that the accuracy of the CNN land surface temperature retrieval algorithm is high.(3)Based on the retrieval results of long-term sequence of soil moisture and land surface temperature,the spatio-temporal changes of key parameters of water conservation in China from 2003to 2018 were analyzed.Trend analysis was used to analyze the temporal changes of soil moisture and surface temperature in six major regions of China.In terms of soil moisture,the change of soil moisture time in China in the past 15 years is relatively large,and the soil moisture in summer is higher than other seasons.From the perspective of spatial distribution,northern Hubei,Henan,Anhui,etc.north of the Yangtze River in China where the soil moisture showed a significant increase in the four seasons,indicating that the surface soil in these areas has obvious humidification trend in the past 15 years,and the water conservation degree has gradually increased.In terms of land surface temperature,the average surface temperature in China has changed slightly in the past 15 years.From a spatial perspective,land surface temperature in spring is generally decreasing.Contrary to spring,most of the surface temperature in winter shows a slight upward trend in China.