基于高光谱的三江源区土壤有机质含量反演

土壤有机质（SOM）是指土壤中各种含碳有机化合物的总称，其动态变化不仅影响农业生态系统的稳定，而且与大气圈和生物圈的碳循环密切相关，对土壤有机碳的大规模快速监测和碳储量核算具有重要意义。本研究于2017年、2018年7月在三江源区野外采集了145个土壤样品，检测了土壤光谱信息。然后将原始光谱反射率数据及其不同数据变换形式下的光谱分别与土壤有机质（SOM）含量进行相关分析，并选取了特征波段，此外利用偏最小二乘回归（PLSR）、支持向量机（SVM）和随机森林（RF）模型对三江源区SOM含量进行建模估算。结果表明，不同深度土壤有机质含量差异明显，且呈逐层下降趋势。而三种建模方法的检验精度分别为:RF > SVM > PLSR，其中RF和一阶微分（FD）组合模拟最好（建模集和验证集的R2、RMSE分别为0.9678、8.9132和0.7841、20.9787）。对于三江源土壤有机质含量反演，不同模型的最佳数据变换方法不同。本研究成果能为后续的高光谱遥感反演提供理论支撑，从而实现三江源区土壤有机质含量的快速检测和实时动态监测。

Hyperspectral Inversion of Soil Organic Matter Content in the Three-Rivers Source Region

Soil organic matter (SOM) refers to the general term of all kinds of carbon (C)-containing organic compounds in soil. Its dynamic change not only affects the stability of agricultural ecosystem, but also is closely related to the C cycle of atmosphere and biosphere. It is of great significance to the large-scale monitoring of soil organic C content and C storage. In this study, 145 soil samples were collected from the field of the Three-Rivers Source region in July 2017 and 2018 to detect the spectral information of soil. Then, the correlations between original spectral reflectance data and the spectrum under different data transformation forms of SOM were carried out, and the characteristic bands were selected. In addition, the partial least square regression (PLSR), support vector machine (SVM) and random forest (RF) models were used to simulate and estimate the content of SOM in the Three-Rivers Source region. The results showed that the content of SOM was significantly different among different soil depths, and showed a downward trend with soil depths. The test accuracy of the three modeling methods was decreased in the order of RF > SVM > PLSR. The combined model of RF and FD (the first-order differential) showed the best simulation accuracy (R2 and RMSE of modeling set were 0.9678 and 8.9132 and those of verification set were 0.7841 and 20.9787, respectively). For the inversion of soil organic matter content in the three river source, the best data transformation methods of different models are different. The results of this study could provide a theoretical support for the subsequent hyperspectral remote sensing inversion, so as to realize the rapid detection and real-time dynamic monitoring of SOM content in the Three-Rivers Source region.