基于PROSPECT模型与高光谱数据的叶片干物质含量估算研究

Study on Estimation of Leaf Dry Matter Content Based on PROSPECT Model and Hyperspectral Data

摘要: 本研究利用统计模型建立了不同光谱尺度下四种植被指数与叶片干物质含量(LDMC)之间的定量关系，这为干物质含量的精准估算提供了相应的理论支持，对监测植物的长势和健康状况也具有一定的指示意义。实验数据由PROSPECT-5模型模拟数据，LOPEX和ANGERS实测数据构成，其中模拟数据用于寻找敏感波段并据此对实测数据进行重采样，处理后的实测数据一半用于建立模型，剩余的另一半则用来验证模型的精度。研究方法是先构建LDMC敏感植被指数，接着分别对其进行光谱尺度敏感性分析，最后可以构建出20个反演模型。总的来说，波段宽度对估测LDMC的影响较小，主要是植被指数的选取会对精度产生一定的影响。为确保结论的可靠性，本研究把未参与建模的实测数据组合成三类，从决定系数R²和均方根误差RMSE两个角度分别对每个反演方程进行精度验证。相比之下用LOPEX和ANGERS数据集验证精度最具说服力，其结果表明：NDVI和RVI所构建的方程精度比较高，拟合度R²可达0.66左右；RMSE较低，约为0.0021，DVI模型的验证结果最不理想，R²只分布在(0.13, 0.25)区间，且RMSE最高，为0.0037。最终的结论是RVI (1688, 1718)模型在不同光谱尺度下估测LDMC的精度水平较高，因此该植被指数具有较好的应用前景。其中10nm为RVI模型的精度最佳光谱尺度，R²是0.6593，RMSE仅为0.002。

Abstract: In this study, a statistical model was used to establish the quantitative relationship between the four vegetation indices and the leaf dry matter content (LDMC) at different spectral scales. This provides corresponding theoretical support for the accurate estimation of LDMC, and monitors the growth and health of plants. It also has certain directive significance. The experimental data is composed of PROSPECT-5 model simulation data and LOPEX and ANGERS measured data. The simulated data is used to find the sensitive band and resample the measured data accordingly. Half of the processed measured data is used to build the model, and the remaining half is used to verify the accuracy of the model. The research method is to first construct the LDMC sensitive vegetation index, then separately analyze the spectral scale sensitivity, and finally construct 20 inversion models. In general, the band width has little effect on the estimated LDMC, mainly because the selection of the vegetation index will have a certain impact on the accuracy. In order to ensure the reliability of the conclusions, this study combined the measured data that did not participate in the modeling into three categories to verify the accuracy of each inversion equation from the perspective of the determination coefficient R² and the root mean square error RMSE. In contrast, the accuracy of using LOPEX and ANGERS data sets to verify accuracy is the most convincing. The results show that the equations constructed by NDVI and RVI have higher accuracy and a fitting degree R² of about 0.66; the RMSE is lower, about 0.0021. The verification results of the DVI model are the most unsatisfactory, R² is only distributed in the interval (0.13, 0.25), and the RMSE is the highest, which is 0.0037. The final conclusion is that the RVI_{(1688, 1718)} model has a higher level of accuracy in estimating LDMC at different spectral scales, so the vegetation index has better application prospects. Among them, 10 nm is the best spectral scale for the accuracy of the RVI model, R² is 0.6593, and the RMSE is only 0.002.