基于多尺度SNV-CWT特征的黑土有机质、水分、总铁及pH值估测

田间土壤属性复杂且随时间变化，快速精准地获得多种土壤理化指标数据对指导精细农业操作具有重要意义。为避免土壤水分带来的干扰，基于光谱技术的土壤成分含量预测需在土壤样本干燥的情况下进行光谱测量，然而土壤水分同样是指导农业生产的重要指标。为同时预测黑土区土壤有机质（SOM）、水分（SMC）、总铁（Fe）和pH值，提出测量湿土土壤样本的可见-近红外光谱，并采用标准正态变量变换（SNV）-连续小波变换（CWT）法分解光谱反射率，逐样本进行SNV后，以Mexh为小波基函数进行10个尺度（21，22，…，210）的分解，并与常用光谱处理方法进行对比，包括高斯滤波（GS）、一阶导数（FD）、连续统去除（CR），数学变换等7种方法。将74个样本数据划分为两组，其中50个作为建模集，24个作为验证集。经SNV-CWT变换后，每个尺度的小波系数与每个目标变量间置信度小于0.05的波段作为随机森林（RF）预测模型的输入变量，以各尺度验证模型精度为标准确定每个预测目标的最佳分解尺度；通过计算最佳尺度小波系数与土壤成分间的皮尔森相关系数（PCC），基于模型的相关系数（MBC）和灰色关联度（GRD），判断各属性的特征波段，且分别以三种相关系数作为指标，以过滤式筛选法建立不同属性的RF估测模型。结果表明：与7种常用的处理方法相比，SNV-CWT分解后四种土壤成分的预测精度均有提高，SOM，SMC，Fe和pH对应的最佳分解尺度分别为7，8，1和10。在以多维特征作为输入变量的情况下，SOM与SMC的验证模型决定系数（R2）即可达到0.90和0.93。三种分析方法中以MBC计算的相关系数为波段筛选指标建立的模型精度最佳，其中SOM与SMC的R2均为0.94，且Fe（R2=0.67，Mse=0.01%，RPD=1.76）与pH（R2=0.80，Mse=0.1，RPD=2.24）的模型精度具有大幅度提高，可应用于多种土壤理化指标数据的提取与监测。

Estimation of Organic Matter,Moisture, Total Iron and pH From Back Soil Based on Multi Scales SNV-CWT Transformation

Soil composition is complex and varied. Predicting the contents of soil propertiesfast and efficiently is important for precision agriculture. Spectra are usually measured on dried soil samples. However, soil moisture is an important indicator for the guidance of agriculture activities. In order to predict the soil organic matter (SOM), soil moisture content (SMC), total iron (Fe) and pH value, we propose to measurement VIS-NIR spectra directly on wet samples and use Standard normal variable (SNV)-Continuous wavelet transform (CWT) method on spectra. CWT method uses Mexh as wavelet filter and 10 scales after SNV on each spectrum. Seven common methods, including Gauss filter (GS), First derivative (FD), Continuous removal (CR), and Mathematical transform (Log(1/R)) et al were used as comparisons. All of 74 samples were divided into 50 and 24, for calibrated and validation datasets. On the coefficients of each scale after SNV-CWT, wavebands that passed 0.05 significance level were selected as RF input variables. The optimal scale for each property was confirmed based on the statistical indicators of validation models. Then the Pearson correlation coefficients (PCC), Model based coefficients (MBC) and Grey relation degree (GRD) between each property and wavelet coefficients were calculated on the optimal scales. Models were estimated by the filter screening method based on the correlation coefficients calculated by the three methods. Results showed that, accuracies of all properties were improved after SNV-CWT comparing to the 7 commonly methods. The optimal transformation scales were 7, 8, 1 and 10, corresponding to SOM, SMC, Fe and pH respectively. When taking high dimension features as input variables, the Coefficient of Determination (R2) was reached to 0.90 and 0.93. The best analysis method was MBC. Because the models performed best when wavebands for the models were selected using MBC as a screening method, the R2 of SOM and SMC was 0.94 and the accuracies of Fe (R2=0.67, Mse=0.01%, RPD=1.76) and pH (R2=0.80, Mse=0.1, RPD=2.24) were greatly improved, methods can be used for extracting and monitoring multi soil properties.