植被电磁传播及散射特性的机器学习

The project focuses on two scientific problems, namely 1) advancement of the understanding of electromagnetic wave propagation and interation with vegetation canopy; and 2) capability of machine learning to extract and characterize wave propagation and scattering mechanism, to study and improve the scattering model of vegetation, to evaluate with high accuracy wave propagation and scattering descriptors such as the extincition coefficient, effective wave number, covariance matrix and coherent matrix, to investigate the best structure of machine learning and learning effect, to conduct a sensitivity analysis of geophysical parameters .so that the main factors affecting wave propagation and scattering can be revealed, and consuquently the assumptions underlying the analytical models can be critically assessed, modified, and refined. Through the study of this project, we will deepen our understanding of the interaction between electromagnetic wave and vegetation, clarify the learning capability machine learning of wave propagation and scattering descriptors, and explore the potenial new approach of microwave scattering mechanism.

本项目围绕两个科学问题，即1）电磁波在植被中的传播交互及散射的深化认识；2）机器学习对波传播及散射机制的特征提取及表征能力，深入研究和改进植被散射模型，准确计算消光系数、等效波数、协方差阵和相干矩阵等波传播与散射描述量，研究机器学习的最佳结构和学习效果，进行地物参数敏感性分析，有效揭示波传播和散射的主要影响因素和机制，从而对解析模型建立中的假设进行评估、修正和完善。通过本项目的研究，加深对电磁波与植被交互的认识，明晰机器学习对波传播与散射描述量的学习能力，探讨微波散射机理研究的新路径与新方法。

在本课题支持下，建立了四种深度神经网络模型以学习柱体的极化双站散射行为，精度高，适用范围广，满足物理约束（能量守恒、互易定理等），有助于准确计算各要素随机分布条件下的协方差矩阵，推动对Durden-Freeman框架下极化分解理论的完善；建立了基于物理信息深度神经网络的逆散射DNN，显示了不同噪声条件下参数反演的鲁棒性；建立了多柱体极化双站散射模型，明确了平均柱体间距对多体耦合效应的影响，以及耦合情况下传播常数和衰减与非耦合情况下情形的差异；改进了扩展先进积分方程模型，显示了对不同粗糙地表条件下的后向及双站散射行为的良好预测。在知名国际期刊（IEEE T-GRS，IEEE GRSL)上已发表5篇标注论文，已接收2篇，另有6篇在准备。通过本项目及后续研究，有望为建立复杂植被覆盖地表场景的高精度高效散射模型、提高极化分解精度、解决极化分解的负功率问题、以及高分情形下极化分解理论的完善提供新的思路和方法。

内容获取失败，请点击重试