基于偏振反射模型和随机森林回归的叶片氮含量反演

叶片氮含量极大程度上影响植被生物化学过程，有重要的研究意义。利用机载高光谱数据反演叶片氮含量在农业遥感领域有广泛应用，但其反演精度不能完全满足精细农业的需要，有一定提升空间。叶片氮含量遥感反演精度受机理误差和算法误差的影响，机理误差主要来源于叶片表面反射。传感器探测到的反射辐射既包含叶片内部多次散射，又包含叶片表面镜面反射部分，只有前者是携带叶片内部生化组分（如氮含量）信息的，由于后者是入射光在叶表蜡质层发生的直接反射，因此该部分并不携带叶片内部信息。根据菲涅尔定律，叶表镜面反射是部分偏振的，而内部散射是非偏振的，因而通过偏振反射建模可部分去除叶表镜面反射影响，以消除机理误差。算法误差主要来源于不同氮含量反演算法对于高光谱数据挖掘能力的差别。比较了偏最小二乘法、主成分回归、支持向量机、K-近邻算法和随机森林回归在高光谱叶片氮含量反演中的表现，在调整算法参数之后，选择使用随机森林回归算法以减少高光谱反演算法误差。以常绿针叶林、落叶阔叶林和针阔混交林为研究对象，利用多角度偏振卫星POLDER/PARASOL的多光谱数据库构建二向偏振反射模型，用以模拟和分析研究区森林的偏振反射率；从HySpex传感器系统获取的机载高光谱数据中去除偏振反射率带来的光谱机理误差，以实现叶片氮含量的精确反演。以均方根误差为主要指标评估精度变化可获得以下结论：在高光谱叶片氮含量反演中，消除偏振反射率带来的机理误差后，各算法反演精度均有提升，平均提升了4.244%。其中，随机森林回归可以最大程度减小反演算法误差（可决系数达到0.803，均方根误差达到0.252），且对光谱偏振信息最为敏感，去除偏振后精度提高了13.103%。相比于广泛使用的偏最小二乘算法，去除光谱机理误差并减小反演算法误差后，叶片氮含量反演精度整体提高了32.440%。该研究实现了基于机载高光谱数据的叶片氮含量精确反演，证明了在叶片氮含量反演中去除偏振反射率的必要性，体现了在高光谱氮含量反演中随机森林算法的应用潜力。

Leaf Nitrogen Concentration Retrieval Based on Polarization Reflectance Model and Random Forest Regression

Leaf nitrogen concentration is of great significance in the vegetation biochemistry process. Airborne hyperspectral data is widely utilized to retrieve leaf nitrogen concentration. Since the current algorithms cannot completely fulfill the accuracy requirement of precision agriculture, it is urgent to improve the retrieval accuracy of leaf nitrogen concentration. The accuracy of leaf nitrogen concentration retrieval is restricted by principle error and algorithm error. The principle error is generated in the process of specular reflection at the leaf surface. The radiant energy detected by sensors consists of a specular components and multiple scattering components. Solely the multiple scattering component carries vegetation biochemistry information (leaf nitrogen concentration, for instance). The specular component represents the energy reflected directly at the foliar wax layer, thus carries no inner information of the leaf. Based on the Fresnel formula, the specular component is partially polarized, and the multiple scattering component is unpolarized. Therefore, the principle error can be eliminated by the specular reflectance estimate, particularly with the aid of polarization reflectance modelling. The algorithm error is derived from the difference of airborne hyperspectral data mining capability between different algorithms. The performance of Partial Least Squares Regression, Principal Component Regression, Support Vector Regression, K-Nearest Neighbor Regression and Random Forest Regression are systematically compared in this research, and ultimately Random Forest Regression is chosen to reduce the algorithm error. In this research, in order to estimate the polarization reflectance of broadleaf and needle vegetation, multispectral data gained by POLDER/PARASOL satellite (equipped with multi-angle polarization sensors) are used to establish Bidirectional Polarization Distribution Function model. Hyperspectral data gained by the HySpex sensor system is used to conduct high-precision retrieval of leaf nitrogen concentration. Root Mean Square Error is taken as a major evaluation index. The conclusion is: After eliminating polarization reflectance in hyperspectral data, an average accuracy improvement of 4.244% is achieved among the above algorithms. Random Forest Regression is rather competitive by reaching 13.103% improvement in accuracy (RSQ 0.803, RMSE 0.252), which indicates that Random Forest is sensitive to polarization information. Compared to the basic method (Partial Least Squares Regression), the accuracy is improved by 32.440% after eliminating principle error and reducing algorithm error. In our research, the high-accuracy retrieval of leaf nitrogen concentration is realized, proving the significance of eliminating polarization reflectance and indicates the potential of random forest regression in hyperspectral remote sensing retrieval.