土壤高光谱在采集过程中难以避免噪声干扰, 造成高光谱数据信噪比较低, 影响土壤有机质含量估测精度。 尝试探究小波能量特征方法, 降低高光谱噪声, 提升土壤有机质含量高光谱估测模型性能。 选取湖北省潜江市运粮湖管理区为试验区, 于2016年9月采集80份深度为0~20 cm的水稻土样本; 土壤样本经风干、 碾磨、 过筛等一系列处理后, 在实验室内采集样本光谱, 并通过重铬酸钾-外加热法测定土壤有机质含量; 利用浓度梯度法, 将总体样本集(80个样本)划分为建模集(54个样本)和验证集(26个样本); 以mexh为小波基函数进行连续小波变换(continuous wavelet transformation), 将土壤高光谱转换为10个分解尺度的小波系数(wavelet coefficients); 逐尺度计算小波系数的均方根作为小波能量特征(energy features), 将10个尺度的小波能量特征组成小波能量特征向量(energy features vector); 逐尺度逐波长计算小波系数与有机质含量的相关系数, 将达到极显著水平(p<0.01)的小波系数作为敏感小波系数(sensitive wavelet coefficients); 利用主成分分析法(principal component analysis)分别计算土壤高光谱和小波能量特征向量的各主成分载荷, 通过比较两者第一主成分贡献率的高低和两者前三个主成分得分的空间离散程度, 判断小波能量特征转换前后建模自变量的主成分信息变化趋势; 基于小波能量特征向量和敏感小波系数分别建立多元线性回归和偏最小二乘回归土壤有机质含量估测模型。 结果表明, 土壤有机质含量越高, 全波段反射率越低, 但不同土样的光谱反射率曲线特征相似, 近红外部分的反射率(780~2 400 nm)高于可见光部分(400~780 nm); 敏感小波系数对应的波长为494, 508, 672, 752, 1 838和2 302 nm; 土壤高光谱与小波能量特征向量的第一主成分贡献率分别为96.28%和99.13%, 小波能量特征向量的前三个主成分散点较土壤高光谱的主成分散点在空间上更为聚集, 表明小波能量特征方法有效减少了噪声影响; 比较全部土壤有机质含量估测模型, 以小波能量特征向量为自变量的多元线性回归模型具有最佳估测精度, 其验证集决定系数(R2)、 相对估测误差(RPD)和均方根误差(RMSE)分别为0.77, 1.82和0.82。 因此, 小波能量特征方法既能够提高数据的信噪比, 提升土壤有机质含量的估测精度, 又实现了土壤高光谱数据降维, 降低了模型复杂度, 可用于土壤有机质含量快速测定和土壤肥力动态监测等研究。

There is no silver-bullet solution of eliminating noise during the acquisition process of soil hyperspectral. As the noise interference, the observations of soil spectra are in low signal-to-noise ratio, which affects the estimation accuracy of soil organic matter content. This paper attempts to adopt the wavelet energy features method to reduce the noise in soil hyperspectral and improve the estimation accuracy of soil organic matter content. The Yunlianghu Farm of Qianjiang City, Hubei Province, located in the hinterland of Jianghan Plain, was selected as the experimental area, and 80 samples of paddy soil with a depth of 0~20 cm were collectedin September 2016. After pretreatment (air drying, milling, sieving), soil sample spectral reflectance and determine soil organic matter contentwere collected in the laboratory. The concentration gradient method was employed to divide the whole sample set (80 samples) into a calibration set (54 samples) and a validation set (26 samples). Continuous wavelet transformation was performed using mexh as a wavelet basis function, transforming the soil hyperspectral into sensitive wavelet coefficients of 10 decomposition scales. Then the root mean square of the wavelet coefficients was calculated scale by scaleto define wavelet energy features, and the wavelet energy features vector was determined by the wavelet energy features. The correlation coefficients between the wavelet coefficients and the organic matter content were calculated scale by scale and wavelength by wavelength, and the wavelet coefficient which reaches the extremely significant level (p<0.01) was defined as the sensitive wavelet coefficients. Principal component analysis was conducted to calculate the principal component loads of soil hyperspectral and wavelet energy features vector, respectively. The trend of principal component information of modeled independent variables before and after wavelet energy features transformation would be judged from the difference between the first principal component contribution rate and the spatial dispersion of the first three principal component scores degree. Moreover, regression models were established based on wavelet energy features vector and sensitive wavelet coefficients, respectively, to estimate soil organic matter content. The results showed that with the increase of soil organic matter content, the full-band reflectance decreased, but the spectral reflectance curves of different soil samples were similar, and the reflectance in the near-infrared bands (780~2 400 nm) was higher than that in the visible bands (400~780 nm). The sensitive wavelet coefficients corresponded to wavelengths of 494, 508, 672, 752, 1 838, and 2 302 nm. The first principal component contribution rates of soil hyperspectral and wavelet energy features vector were 96.28% and 99.13%, respectively. The first three principal component scatter points of wavelet energy features vector were more spatially aggregated than those of soil hyperspectral, which demonstrated that the wavelet energy features method effectively reduces the influence of noise. Comparing the estimation models of soil organic matter content, the multivariate linear regression model adopting wavelet energy features vector as the independent variable had the highest estimation accuracy, whose determination coefficients (R2), relative estimate deviation (RPD), and the root mean squared error (RMSE) of validation set were 0.77, 1.82, and 0.82, respectively. Therefore, the wavelet energy features method which is proved to raise the signal-to-noise ratio of the data without adding to the complexity, could improve the estimation accuracy of soil organic matter and realize the dimensional reduction of soil hyperspectral data. This method can be applied to studies like on-the-go soil properties measurement and soil quality monitoring.