Soil Moisture Data Recontruction Based On Deep-Learning And Matrix Completion

Soil moisture is an important physical quantity in land-air interaction,which plays an important role in the research and practical application of meteorology,hydraulics,agriculture,forestry,climate and many other disciplines,and is an important parameter affecting surface evapotranspiration,runoff,surface albedo,surface emissivity,and surface sensible heat and latent heat flux.At present,the observed soil moisture is quite scarce,and its applicability is greatly limited.Therefore,to obtain large-scale and high-precision soil moisture data is one of the research hotspots in the field of meteorology.Early soil moisture reconstruction methods were mainly based on spatial interpolation.However,the spatial interpolation method only focuses on the local characteristics of soil moisture,ignoring its overall correlation.In recent years,the emerging Matrix Completion(MC)method can realize the reconstruction of largescale low-rank or approximate low-rank matrices lacking a large number of values.Soil moisture data meets the conditions of approximate low-rank,and can capture data features better than traditional interpolation algorithms.Therefore,this paper applies the matrix filling algorithm to the reconstruction of soil moisture,and its reconstruction accuracy is significantly improved compared with the traditional interpolation algorithm.Principal Component Analysis(PCA)and Autoencoder(AE)extract the characteristics of data features.This paper innovatively proposes the PCA-MC algorithm and AE-MC network.By extracting the main features of low-rank matrix,To achieve matrix dimension reduction and improve algorithm performance.Through comparative analysis,compared with using matrix filling algorithm alone,PCA-MC and AE-MC have significant advantages in reconstruction accuracy and time loss.Compared with using MC algorithm alone,the average deviation of PCA-MC reconstruction algorithm is reduced by 27.1%,the average root mean square error is reduced by 4.60%,and the average maximum error is reduced by 12.8%.Compared with MC algorithm alone,the mean error of AE-MC reconstruction results is reduced by 32.3%,the mean root error is reduced by 12.0%,and the mean maximum error is reduced by 19.3%.Therefore,PCA-MC and AE-MC can better reconstruct the incomplete matrix,which has new progress for the research of soil moisture interpolation algorithm.The contents of the research on soil moisture data reconstruction in this paper are as follows:(1)First,obtain ERA5 reanalysis data,introduce the characteristics of ERA5 reanalysis soil moisture data,and establish soil moisture data set.(2)The traditional spatial interpolation algorithm and the classical matrix filling algorithm are introduced,and the advantages of matrix filling algorithm in soil moisture data reconstruction are compared and analyzed.(3)The introduction of PCA algorithm,PCA algorithm can extract the main features of data,achieve dimensionality reduction,put forward the reconstruction principle based on PCAMC algorithm,and analyze its advantages in data reconstruction.An algorithm model suitable for EAR5 reanalysis of soil moisture data was constructed for data reconstruction.(4)The AE network was introduced to obtain the overall characteristics of soil moisture data,and the soil moisture data reconstruction based on AE-MC network was proposed by using the noise reduction and anti-noise properties of the noise reduction autoencoder.