土壤有机质含量是反映土壤肥力的重要指标, 对其进行动态监测是实施精准农业的重要措施。 近年来, 众多学者尝试采用土壤近地传感（proximal soil sensing）, 尤其是近地高光谱技术, 在田间和实验室获取不同形态土壤的高光谱数据, 不断引入新方法建立适用于不同地域和不同土壤类型的有机质含量的反演模型。 该研究在实验室内利用ASD FS3采集了土壤高光谱数据, 采用“重铬酸钾-外加热法”测得了土壤有机质含量； 分析了土壤原始光谱反射率（R）与有机质含量的相关性, 选取R2>0.15的敏感波段的反射率； 利用CWT对土壤原始光谱反射率（R）、 光谱反射率的连续统去除（CR）进行不同尺度的分解, 分析小波系数与土壤有机质含量的相关性, 选取R2>0.3的敏感波段的小波系数； 利用R选取的波段信息和R-CWT, CR-CWT的选取的小波系数, 分别建立偏最小二乘回归（PLSR）、 BP神经网络（BPNN）、 支持向量机回归（SVMR）三种不同的土壤有机质含量反演模型。 结果表明： 相比R与土壤有机质含量的决定系数R2, R-CWT, CR-CWT变换后得到的小波系数与土壤有机质含量的决定系数R2分别提高了0.15和0.2左右； CR-CWT-SVMR的模型效果最为显著, 预测集的R2和RMSE分别为0.83, 4.02, RPD值为2.48, 具有较高的估测精度, 能够全面稳定地估算土壤有机质含量； CR-CWT-PLSR的模型精度与CR-CWT-BPNN, CR-CWT-SVMR相比虽有一定差距, 但是其计算量要明显小于非线性的BPNN和SVMR方法, 具有模型简单、 运算速度快等特点, 对开发与设计田间传感器具有较大的应用价值。

Soil organic matter content (SOMC) is an important parameter that reflect soil fertility available for crop production, and monitoring of the SOMC dynamically has shown great importance to promote the development of precision agriculture. In recent years, many researchers have tried to use proximal soil sensing, especially using the proximal hyperspectral techniques to acquire different kinds of spectral data under the field and laboratory conditions, and various new algorithms are also introduced to build inversion models to predict SOMC from spectra for different regions and different kinds of soils. In this paper, the hyperspectral reflectance of different soil samples was measured using the ASD FieldSpec3 spectrum analyzer. At the same time, the SOMC of each soil sample was analyzed using potassium dichromate external heating method in the laboratory. The correlation analyses between raw soil spectral reflectance (R) and SOMC were done, and it could select sensitive wavebands reflectance when the determination coefficients (R2) exceeded 0.15. A continuous wavelet transform (CWT) was also performed on R and the continuum removal curves (CR) to generate a wavelet power scalogram in different scales, the correlation analyses were done between wavelet power coefficients and SOMC, and it could select the sensitive wavelet coefficients when the R2 exceeded 0.3. Then, after extracting wavebands reflectance from R and wavelet power coefficients from R-CWT, CR-CWT, the estimation models for SOMC had been successfully built by partial least squares regression (PLSR), BP neural network (BPNN), support vector machine regression (SVMR), respectively. The results showed that, compared to the R2 between SOMC and R, the R2 between SOMC and R-CWT, CR-CWT wavelet coefficients were increased by about 0.15 and 0.2. The CR-CWT-SVMR model was the best, its R2, RMSE and RPD value of validation set were 0.83, 4.02, 2.48, which could estimate SOMC comprehensively and stably. For the CR-CWT-PLSR model, although there was a slight gap in the prediction accuracy with that CR-CWT-BPNN and CR-CWT-SVMR models, it also had its own unique advantages: the model was simple and thus the computation speed was reduced significantly. In the future, the results can provide good potential for field proximal sensing researching.