高光谱成像的土壤剖面水分含量反演及制图

Inversion and Mapping of the Moisture Content in Soil Profiles Based on Hyperspectral Imaging Technology

传统土壤水分的获取方法仅可获得离散的土壤水分点位数据,难以获得剖面上精细且连续的水分含量分布图。研究了野外条件下利用近红外高光谱(882~1709nm)成像反演剖面土壤水分含量(SMC),并实现精细制图的可行性。研究剖面位于江苏省东台市,我们利用近红外高光谱成像仪对剖面进行了5天原位连续观测,共采集了280个土样用于烘干法测定SMC。原始高光谱图像经数字量化值(DN)校正、黑白校正、拼接、几何校正、剪切和掩膜等一系列预处理后,提取各采样点的平均光谱反射率。提取光谱(Raw)经吸光度[LOG10(1/R)],Savitzky-Golay平滑(SG)、一阶微分(FD)、二阶微分(SD)、多元散射校正(MSC)和标准正态变量(SNV)转换后,采用偏最小二乘回归(PLSR)和最小二乘支持向量机(LS-SVM)方法建立SMC预测模型,并对比分析不同光谱预处理方法与建模方法组合条件下SMC的预测精度。结果表明,光谱反射率随SMC增加逐渐降低,不同光谱预处理方法的预测精度有所差异,除MSC方法外,同一光谱预处理方法的LS-SVM模型预测精度均高于PLSR模型,并且基于LOG10(1/R)光谱的LS-SVM模型对SMC预测精度最高,其建模集的决定系数(R2c)和均方根误差(RMSEc)分别为0.96和0.65%,预测集的决定系数(R2p)、均方根误差(RMSEp)和相对分析误差(RPDp)分别为0.88,1.05%和2.88。利用最优模型进行剖面SMC的高空间分辨率精细制图,通过比较SMC反演图中提取的预测值与实测值关系发现预测精度较高(R2:0.85~0.95,RMSE:0.94%~1.02%),且两者在剖面中的变化趋势基本一致,说明SMC反演图不仅能很好地反映出土壤水分在整个剖面中毫米级的含量分布信息,也可反映出同一位置处不同天数间的含量差异。因此,利用近红外高光谱成像结合优化的预测模型,能够实现土壤剖面SMC的定量预测及精细制图,有助于快速、有效监测田间剖面土壤水分状况。

Traditional methods for acquiring soil moisture can only provide discrete point data, which arenot so appropriate for finely and continuously mapping soil mois ture distribution in soil profile. In this paper, the feasibility of predicting and mapping ofsoil moisture content (SMC) in soil profile was studied using ne ar-infrared hyperspectral imaging in the spectral range of 882~1 709 nm. Two soil profiles, located in Dongtai City of Jiangsu Province, were continuously observed in situ for 5 days by the near-infrared hyperspectral imaging system. A total of 280 soil samples were obtained for later SMC measurement by oven-drying method. After a series of preproces on the acquired raw hyperspectral images, including digital number(DN) correction, reflectance correction, mosaicki ng, geometric correction, image clipping and masking, the average spectral reflectance of each sampling point in the corrected hyperspectral images was extra cted for further analysis. Then the extracted spectra (Raw) were preprocessed b y LOG 10 (1/ R ), Savitzky-Golay (SG), first derivative (FD), second d erivative (SD), multiplicative scatter correction (MSC) and standard normal var iate (SNV), and partial least squares regression (PLSR) and least squares support vector machine (LS-SVM) models were developed and comparedfor a selection of optimum prediction model. Results showed that the soil spectral reflectance gra dually decreased with the increase of SMC, and different spectral preprocessing methods had different prediction accuracy. Except for the MSC preprocessing me thod, the prediction accuracy of the LS-SVM model was higher than the PLSR model with the same spectral preprocessing method. The prediction accuracy of the LS-SVM model with LOG 10 (1/R) preprocessed spectra was highest with R^2 c of 0.96 and RMSE c of 0.65% for calibration, and R^2 p of 0 .88, RMSE p of 1.05% and RPD p of 2.88 for prediction. The optimum model was then applied to produce high spatial resolution maps of SMC in profiles. The prediction accuracy was high ( R^2: 0.85~0.95, RMSE: 0.94%~1.02%) by comparing the extracted SMC values from prediction maps with the measured values, and both SMC had the samedistribution tendency in profiles, demonstrating that the SMC prediction mapscould well displaynot only the SMC distribution in prof iles in the millimeter scale, but also the changes of SMC at different location s in the profile between different days. Thus, the near-infrared hyperspectral imaging technology combined with optimized prediction model could provide a new approach to quantitatively predict and map high spatial resolution images of SMC in soil profiles in situ, which could help to rapidlyand effectively monitor soil moisture in profiles in the field.