结合辐射传输模式和机器学习技术发展云与气溶胶的激光雷达快速正反演算法

申请人计划结合辐射传输模式和深度学习技术发展云与气溶胶的激光雷达快速正反演算法。首先，结合散射数值模式、实验室测量数据、非弹性激光雷达观测数据研究云、气溶胶的后向散射特性，并建立一个能够很好地模拟实际激光雷达信号的单粒子散射特性数据库。使用散射特性数据库和蒙特卡罗辐射传输模式建立一套多次散射强度及其影响因子之间的关系的数据集，结合深度学习算法建立一个激光雷达信号模拟的快速正演算法。基于快速模拟算法和卷积神经网络建立一套云与气溶胶的识别、分类和反演算法，并通过强化学习使这套算法的精度在实际应用中得到改进。

We plan to develop a fast algorithm to simulate the lidar signals, and an algorithm to retrieve the particle properties with a combination of Monte Carlo radiative transfer model and deep learning algorithms. First of all, we will further study the backscatter properties of cloud and aerosol particles with models and observations, and build a scattering database that could be used to simulate the observed lidar signals of clouds and aerosols. Subsequently, we will investigate the relationship between the multiple scattering intensity and its impact factors with Monte Carlo radiative transfer model, and train a convolutional neural network to predict the lidar signals. Finally, a retrieval system will be built to identify, classify and retrieval the properties of clouds and aerosols. The performance of the retrieval system is expected to improve via reinforcement learning.