去除水分影响提高土壤有机质含量高光谱估测精度

土壤水分的影响是当前采用光谱分析法预测土壤养分含量的关键问题,该文旨在探索去除土壤水分影响、提高有机质高光谱定量估测精度的方法。首先采用地物光谱仪进行湿土和过筛干土的高光谱测试,并进行一阶导数变换;然后,采用奇异值分解(singular value decomposition,SVD)结合相关分析筛选土壤水分特征光谱,构建去除水分因素的修正系数,形成湿土光谱的校正光谱;最后基于校正前后湿土光谱,应用偏最小二乘(partial least squares,PLS)回归构建土壤有机质含量的估测模型,并对模型进行验证和比较,分析评价校正前后光谱的预测精度。结果显示:按土壤水分含量梯度划分的2组和全部棕壤及褐土土样共4组样本校正后建模决定系数和均方根误差分别为0.85、0.82、0.74、0.76和0.19%、0.20%、0.23%、0.19%,决定系数提高了0.02~0.09,均方根误差降低了0.01~0.03百分点,验证决定系数、均方根误差和相对分析误差分别为0.78、0.77、0.72、0.76,0.21%、0.15%、0.21%、0.15%和2.03、2.02、1.86、1.98,决定系数提高了0.06~0.15,均方根误差除褐土土样提高0.02百分点外,其他样本组降低了0.01~0.08百分点,相对分析误差提高了0.17~0.43,模型决定系数和相对分析误差得到显著提升;尤其对于土壤水分含量变异系数较小的3组土样,模型从待改进级别提高到性能良好级别,对土壤有机质含量具有较好的预测准确性。说明该方法用于去除土壤水分因素影响和提高有机质含量高光谱估测精度的有效性。

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.