土壤光谱重建的湿地土壤有机质含量多光谱反演

土壤有机质是湿地生态系统的重要元素，利用多光谱遥感技术可大尺度、快速获取其含量信息，对保护湿地生态系统具有重要意义。然而，由于不同地物光谱混合给多光谱数据带来光谱畸变，影响湿地土壤有机质含量的反演精度。为了消除不同地物光谱混合，实现湿地土壤有机质含量的准确、实时监测，以闽江鳝鱼滩湿地为研究区，利用线性波谱分解技术对原始影像的像元进行分解，重建土壤光谱，分析原始光谱、重建光谱与土壤有机质含量的相关性后，建立土壤有机质含量的反演模型。结果表明：利用线性波谱分解技术可有效消除原始影像中的植被端元，减少大部分道路及建筑物的反射干扰，重建后的土壤光谱特征曲线更趋近于自然状态下土壤的光谱曲线，重建效果显著；通过两种光谱与土壤有机质含量的相关系数对比，重建光谱更能准确的反映土壤光谱与土壤有机质含量的相关性；运用重建光谱构建土壤有机质含量的反演模型，其预测精度优于基于原始光谱的反演模型，R2和F分别提高0.124和2.223，RMSE则降低0.106，1∶1线检验的预测值与实测值的拟合度更高，模型可行且有效。由此得出结论，利用线性波谱分解技术消除不同地物光谱混合，重建土壤光谱，一定程度上可实现在自然条件下湿地土壤有机质含量的大面积、准确检测，具有较好的实际应用价值。

Inversion of Soil Organic Matter Content in Wetland Using Multispectral Data Based on Soil Spectral Reconstruction

Soil organic matter (SOM) is an important element of wetland ecosystem. Quick and wide monitor of SOM content with multispectral remote sensing technique has the vital significance to protect wetland ecosystem. More previous studies on the estimation of SOM content used hyperspectral analysis, while using multispectral was less. The main reason is that the spectral anomaly of multispectral data caused by spectral mixing of different objects affects the inversion accuracy of SOM content in wetland. Therefore, to avoid the spectral anomaly, this paper took the Shanyutan wetlands of Minjiang River Estuary as a survey region, trying to use Linear Spectral Unmixing Model(LSUM) to separate the pixel of original image and reconstruct the soil spectrum. Then, the correlation analyses between 2 different spectra (the raw spectrum and the reconstructed spectrum) and SOM content were done. Finally, according to correlation results, an inversion model for SOM content was established. The result showed that LSUM can effectively eliminate vegetation endmembers of the original image, reducing the reflection interference of most roads and buildings. The reconstructed spectral characteristic curve was closer to the spectral curve of soil under natural condition. It indicated that the effect of spectral reconstruction was remarkable; Compared to the correlation coefficients between 2 different spectra and SOM content, the reconstruct spectrum was more appropriate for reflecting the correlation between the soil spectrum and soil organic matter in the study area; using the reconstructed spectrum to build the predicting model could obtain more robust prediction accuracies than using the raw spectrum. Its values of R2 and F were increased by 0.124 and 2.223 respectively. And RMSE was reduced by 0.106. Moreover, through the 1∶1 line test, model of the reconstructed spectrum had a better fitting between the predicted and the measured. These results suggested that using LSUM has been proven to be effective in removing the spectral anomaly, ensuring a transferrable model for SOM content under natural condition. The study will provide some practical technology to monitor the SOM content in wetland by multispectral data.