Estimation of soil organic carbon content in farmland based on UAV hyperspectral images:A case study of farmland in the Huangshui River basin
Rapid and accurate estimation and spatial distribution mapping of soil organic carbon content in farmland facilitate the refined management of soil and the development of smart agriculture.This study investigated three typical farmland areas in the Huangshui River basin of Qinghai Province using 296 soil samples and corresponding field in situ spectra collected synchronously.The unmanned aerial vehicle(UAV)with a hyperspectral camera was employed for image acquisition,and the soil samples were tested for spectral acquisition and organic carbon content in the laboratory.The spectral reflectance was transformed into seven different forms,and the main characteristic bands were screened out through correlation analysis.Using multiple linear regression,partial least squares regression,and random forest,the experimental spectra,field in situ spectra,and UAV spectra were modeled,with the accuracy of the models compared.The UAV spectra were corrected using the direct spectral conversion method,and the optimal model of corrected UAV spectra was used for modeling.The model was substituted into the UAV hyperspectral images for the organic carbon content mapping.Finally,the farmland areas meeting the mapping accuracy requirements were analyzed and discussed.The results show that:① The multiple linear regression after logarithmic transformation of UAV hyperspectra failed to estimate the organic carbon content,with a relative percent deviation(RPD)of 1.375.Except for it,the experimental spectra,field in situ spectra,and original spectra of UAV hyperspectra as well as all conversion methods could estimate the organic carbon content,with coefficients of determination(R2)ranging from 0.562 to 0.942,root mean square errors(RMSEs)ranging from 1.713 to 5.211.and RPDs between 1.445 and 4.182;② Among all spectral transformation methods,multiple scatter correction and first-order differential transformation exhibited the highest correlation with the organic carbon content,presenting characteristic bands of 429～449 nm,498～527 nm,830～861 nm,and 869 nm;③ As revealed by the modeling results,the random forest model manifested the highest accuracy,followed by the partial least squares model and the multiple linear regression model in turn.The corrected UAV spectra yielded improved modeling accuracy;④ The inversion accuracy of the three farmland areas all met the mapping requirements,with R2 values above 0.88.Farmland A exhibited the highest average organic carbon content of 28.88 g·kg-1 and an overall uniform spatial distribution.Farmland B manifested average organic carbon content of 13.52 g·kg-1 and a significantly varying spatial distribution.Farmland C displayed the lowest average organic carbon content of 8.54 g·kg-1 and significant differentiation between high and low values.This study can be referenced for the application of UAV hyperspectral remote sensing technology to the field-scale estimation and digital mapping of soil organic carbon content.
万方数据
