Estimate Of Soil Moisture Content Based On Selection Of Variable And Extreme Learning Machine

With the development of hyperspectral technology,due to its high resolution,it had been more and more widely used in the acquisition of soil property information.Soil moisture content was a natural condition such as climate、vegetation、geomorphology、and soil factors.The comprehensive reflection was the basis for the construction of ecosystems and vegetation in arid and semi-arid regions and was an important indicator for measuring land degradation and drought.The rapid monitoring of soil moisture content had important implications for the developme nt of precision agriculture in arid and semi-arid regions.In this experiment,using the Weigan-Kuqain Oasis in Xinjiang as a target area,a total of 39 soil samples were gatherted and brought back to the laboratory,the sampling depth for each sampling point was 0~20 cm.The sample was placed in a thermostatic oven at 105°C for 48 hours to dry it.The dried soil samples were quickly loaded into 45 aluminum box.At the same time,the weight of the soil samples was weighed in the same one thousandth of a metric scale,and then different levels of distilled water were added to the petri dishes.One of them did not add distilled water as a control.the soil moisture in 45 aluminum box was distributed between 0%and 35%(after calculation,the actual sample water content was between 0%and 33.80%),and45 aluminum box were sealed for 48 h,the distilled water added to the Petri dish was fully diffused in the soil and 45 soil test samples were obtained.obtaining soil reflectance spectral curve by ASD Fieldspec3 spectrometer in dark room.After taking the spectrum with a portable spectrometer,quickly remove about 20 g of soil sample into the weighed aluminum box,obta ining moisture content of soil samples by drying method.First,using the db4 function in MATLABR12a,the spectral reflectance of the soil was decomposed into 8 layers,and correlation coefficients between soil moisture content and the spectra of each level was computed.Secondly,through adaptive reweighted sampling(CARS),successive projections algorithm(SPA),and CARS-SPA algorithm,the feature wavelengths of each layer of wavelet feature spectrum within the maximum decomposition scale were selected,finally determined the set of optimal characteristic wavelengths that could reflect the soil moisture content;Thirdly,root mean square error、coefficient of determination a nd relative prediction deviation as evaluation indicators,then comparative analysis was performed to select the best prediction model for predicting soil moisture content.The following conclusions were drawn:(1)With the increase of wavelet decomposition scale,the correlation between soil spectral reflectance and soil moisture content showed a trend of increasing first and then decreasing,and L6 was the most significant band at 0.01 level.In general,the characteristic spectrum of L6 was denoised at the same time,the spectral detail was preserved to the maximum extent,so the maximum decomposition order of the wavelet was 6 order decomposition;(2)In PLSR model,compared with the single-layer feature spectrum after wavelet transform,the prediction models with the highest accuracy in the CARS、SPA and CARS-SPA algorithms were respectively selected,and it was found that the accuracy of the model constructed by the characteristic wavelength obtained by the CARS-SPA algorithm was relatively good.This showed that using the SPA algorithm to perform secondary filtering on the feature bands acquired by the CARS algorithm not only simplifies the model but also improves the model accuracy.O verall,the model constructed by CARS-SPA-FULL-PLSR had the highest accuracy,which indicated that the CARS-SPA secondary screening results of L1~L6 feature spectra after wavelet decomposition were combined as the optimal feature wavelength set,which can increase the accuracy of the model;(3)Spectral reflectance of soil were filtered by coupling the wavelet transform with CARS-SPA algorithm.The optimal feature variable set includes 43 wavelength variables between 400~500,1320~1461,1851~1961 and 2125~2268 nm.(4)On the basis of the set of optimal characteristic wavelengths,different modeling methods were established respectively,and comparative analysis shows that in all soil moisture prediction models,their PRD were all greater than 2.00,and they were sorted by the precision from big to small:ELM>RF>SVM>BP>PLSR,it could be seen that relative to PLSR、BP、SVM,The predictive power of ELM and RF was better,among which CARS-SPA-FULL-ELM has the highest prediction accuracy with RMSEC of 0.0125,R_c~2=0.9858,RMSEP=0.0186,R_p~2=0.9712 and RPD=5.68.This reflected the strong analytical ability of the ELM prediction model for nonlinear problems and the robustness of the model,to provide new ideas for predicting soil moisture content in Weigan-Kuqain Oasis.