Improving estimation precision of soil organic matter content by removing effect of soil moisture from hyperspectra

Soil moisture is a key issue in using spectrum analysis method to predict soil nutrients content. The purpose of this article is to explore a method of removing the effect of soil moisture and improving the hyperspectra estimation precision of soil organic matter (SOM) content. Firstly the soil samples were collected from agricultural fields of the brown soil in Daiyue county and the cinnamon soil in Huantai county, Shandong province, China. The hyperspectra of the moisture and sieved dry soil samples were measured using the ASD FieldSpec 3 and transformed to the first deviation. Because the soil moisture content and its coefficient of variation (CV) of the brown soil samples was relatively high, the brown soil samples were divided into two groups, additionally the all brown soil samples and the cinnamon soil samples, here there were four-group soil samples. Secondly, based on the difference between the moisture and dry spectra, the characteristic spectra of soil moisture were selected by singular value decomposition (SVD) in combination with correlation analysis, then the correcting coefficients of removing moisture factor from soil hyperspectra were built to reconstruct the corrected spectra of the wet samples. Finally the estimation models of the soil organic matter content were built using the partial least squares (PLS) regression based on the uncorrected and corrected spectra of the wet samples. The results indicated that using singular value decomposition to correct the moisture spectra could partly reduce the correlation coefficients between the soil moisture content and the hyperspectra in most range of spectra, and for the four-group soil samples including two for each brown soil grouped by the soil moisture content gradient, all brown soil and cinnamon soil, the coefficient of determination (R2) and relative prediction deviation (RPD) of models based on the corrected spectra were improved signally with the calibration R2 of 0.85, 0.82, 0.74 and 0.76 (an increase of 0.02-0.09), and the calibration root mean squares error (RMSE) of 0.19%, 0.20%, 0.23% and 0.19% (reduce of 0.01%-0.03%), the validation R2 of 0.78, 0.77, 0.72, and 0.76 (an increase of 0.06 and 0.15) the validation RMSE of 0.21%, 0.15%, 0.21% and 0.15% (reduce of 0.01%-0.08%) except for the cinnamon soil samples (increased 0.02 percentage), the validation RPD of 2.03, 2.02, 1.86 and 1.98 (an increase of 0.17-0.43), especially for the three-group samples with the smaller CV in soil moisture content. The models reached the good performance from needing improvement and could achieve better prediction accuracy of the soil organic matter content. Therefore, the experiments indicated that the method was effective to remove the soil moisture influence and improve the prediction accuracy of the soil organic matter content from hyperspectra. In addition, in order to achieve the better prediction accuracy, the soil samples should be grouped by the soil moisture content to reduce the dispersion degree of the soil moisture.