土壤有机碳含量是土壤肥力与土壤质量的主要决定因素, 与土壤生产力密切相关。 采用高光谱模型估算土壤有机碳含量成为了解土壤肥力的重要方法。 利用高光谱分析技术结合机器算法实现快速、 高精度的估算土壤有机碳含量, 对土壤肥力的可持续利用至关重要。 根据实测的土壤有机碳含量及其高光谱反射率数据, 运用Savitzky Golay方法对光谱波段进行平滑去噪, 采用连续投影算法（SPA）、 遗传算法（GA）对原始光谱及其5种不同数学变换光谱分别进行特征波段的筛选, 并基于随机森林（RF）方法构建土壤有机碳含量的高光谱估算模型。 为进一步降低模型的复杂度, 将SPA算法与GA算法相结合, 寻找最佳特征参数, 以提升土壤有机碳含量特征波段的识别率和可信度。 结果表明: （1）在原始光谱中, 基于GA算法筛选SOC含量的高光谱响应波段主要集中在350~410、 827~928、 997~1 064、 1 201~1 234、 1 541~1 574、 1 667~1 710、 2 153~2 186和2 357~2 707 nm; 当RMSE为6.09时, SPA算法筛选了11个特征变量。 （2）基于GA算法筛选特征波段时, 原始光谱R、 标准正态变量（SNV）、 多元散射校正（MSC）、 一阶微分（FD）、 对数的倒数（RL）与连续统去除（CR）的维数分别降低到407、 697、 668、 667、 493、 784维, 占全光谱波段的18.93%~36.47%; 基于GA-SPA算法筛选后, 6种光谱变量的维度介于8~17维, RMSE介于4.53~6.30。 （3）在一阶微分光谱形式下, 基于GA-SPA算法挑选的12个特征变量所构建的RF模型预测效果最好, 模型的建模集R2c为0.78, RMSEc为5.48, 验证集R2p为0.82, RMSEp为4.50, RPD为2.18。 研究表明, 光谱一阶微分可以增强土壤的光谱信息, GA算法结合SPA算法寻找光谱特征变量, 既简化了估算模型的复杂度, 又提高了估算模型的精度, 基于遗传算法—连续投影算法的高光谱模型具有较高的估算能力。

Soil organic carbon content was a major determinant of soil fertility and soil quality and was closely related to soil productivity. The estimation of soil organic carbon content using hyperspectral models has become an important method of understanding soil fertility. Using hyperspectral analysis combined with machine algorithms to achieve rapid and highly accurate estimation of soil organic carbon contents was essential for the sustainable use of soil fertility. Using the measured soil organic carbon content and its hyperspectral reflectance data as the research object, we applied the Savitzky Golay method to smooth and demise the spectral bands, used successive projection algorithm (SPA) and genetic algorithm (GA) to screen the original spectra and its five different mathematical transformed spectra respectively for the characteristic bands, and constructed the random forest (RF) method based on the soil organic carbon content. The hyperspectral estimation model of soil organic carbon content was constructed using the random forest (RF) method. The SPA algorithm was combined with the GA algorithm to find the optimal feature parameters to improve the recognition rate and confidence in the SOC feature bands. The results showed that in the original spectrum, the hyperspectral response bands based on the GA algorithm to screen SOC content were mainly concentrated on 350~410, 827~928, 997~1 064, 1 201~1 234, 1 541~1 574, 1 667~1 710, 2 153~2 186, 2 357~2 707 nm. When the RMSE was 6.09, 11 characteristic variables were screened by the SPA algorithm. The dimension of the original spectrum, standard normal variables (SNV), multiple scattering corrections (MSC), first-order differential (FD), logarithmic reciprocal (RL) and continuum removal (CR) were reduced to 407, 697, 668, 667, 493 and 784 dimensions respectively, accounting for 18.93%~36.47% of the full spectral band when filtering the characteristic bands based of the GA algorithm. After screening based on the GA-SPA algorithm, the dimensions of the six spectral variables ranged from 8 to 17 dimensions, and the RMSE ranged from 4.53 to 6.30. In the first-order differential spectral form, the RF model constructed based on 12 feature variables selected by the GA-SPA algorithm predicted the best results from a modeling set R2c of 0.78 and RMSEc of 5.48, a validation set R2p of 0.82, RMSEp of 4.50, and RPD of 2.18. It was shown that the first-order spectral differentiation could enhance the spectral information about soil, the GA algorithm combined with the SPA algorithm to find the spectral feature variables simplifies the complexity. It improves the accuracy of the estimation model, and the hyper spectral model based on the genetic algorithm-continuous projection algorithm has a high estimation capability.