| The hydrological cycle is an important component of energy and material flow in nature.It promotes the formation and evolution of the natural environment and affects the development of human society.Soil moisture is a key parameter in water and energy cycle between the exchange of the atmosphere and land surface,and in the research of hydrological cycle.The study of soil moisture measurement methods is of great significance to monitoring surface plant evapotranspiration,local climate change,drought and flood disasters,and simulating watershed hydrological processes.The traditional measurements in soil moisture monitoring have large observation time intervals,sparse observation points,and small observation area.Remote sensing inversion methods,compared with the traditional one,have high resolution,wide coverage,and strong timeliness,which is gradually becoming a hot research topic.This paper takes the upstream of Heihe River Basin as the research area.The MODIS daily surface reflectance and temperature remote sensing data were used to construct the temperature vegetation drought index model,after image noise filtering,cloud detection and other preprocessing processes.At the same time,the vegetation index features and the original reflectivity features were used to construct soil moisture inversion models based on BP neural network,support vector regression and kernel ridge regression using both remote sensing data and surface observation data.The inversion accuracy of the two features on the three machine learning models is compared and analyzed.The main results and conclusions of the research are as follows:(1)An improvedσfiltering method is proposed for the speckle noise problem of MODIS remote sensing images.The filtering method considers the pixels in the local filtering window obeying the Gaussian normal distribution,and noise correction detected only when the pixel value outside the interval(μ-2σ,μ(10)2σ)is performed.The results show that the improvedσfilter preserves the details of the image more completely,and effectively filters the strong noise points that the traditionalσfilter fails to.(2)To improve the poor fitting accuracy of dry and wet edges in the inversion method of temperature vegetation drought index,this paper designs a noise reduction method based on the probability of point spacing.The method counts the number of neighboring points of a certain data point(7)VI_i,LST_i~*(8)in the feature space.When the number of neighboring points is less than a certain threshold m,the point is considered as a noise point.The results show that the contours of the scattered point obviously clusters in the feature space processed by the noise reduction method,and the fitting accuracy of dry and wet edges is not easily interfered by noise points.(3)The inversion accuracy of the two input features on the three machine learning models is compared and analyzed.Taking land surface reflectance and land surface temperature as the first group of input features,and vegetation index and land surface temperature as the second group of input features,the soil moisture inversion models based on BP neural network,kernel ridge regression and support vector regression were constructed respectively.The results show that the inversion accuracy of the first group of input features is better than that of the second input features,and the inversion accuracy of the kernel ridge regression is better than the other two machine learning models. |
