New Statistical Methods for Computer-Assisted Inversion with Applications to Satellite Remote Sensing

将基于卫星的遥感数据转换为有用的气候和地球物理信息需要求解辐射传输方程。由于需要在整个星球的尺度上处理数据，目前用于近似求解这些复杂方程的方法有限，并且没有系统地利用多传感器测量、地面测量和正在发挥作用的时间动力学。与此同时，世界各国越来越多地求助于遥感数据，以应对与气候变化和环境退化相关的挑战。为了准确地告知利益攸关方，需要更好的统计模型，特别是对地面测量有限的发展中国家区域进行成像。这项研究的目标是开发新一代统计方法，在更局部的水平上求解卫星遥感数据处理的核心方程。将卫星遥感数据转换为有用的气候和地球物理信息需要解决一些非常重要的辐射传输反问题。该项目旨在开发一个贝叶斯框架，该框架将展开深度学习模型的算法和前向计算机代码结合到一个反演图中。将开发一个转移学习和一个强化学习框架，以将在计算机上学习的反演地图与地面测量结合起来，根据分布失配进行调整，并保持反演程序随时间的准确性，即使卫星数据的分布随着时间的变化也是如此。该项目还将在理论层面上有助于从统计上更深入地理解强化学习和算法展开模型。该项目旨在改善全球遥感数据的分析，将其应用于气候变化、遥感、机器学习和统计学等学科。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The transformation of satellite-based remotely sensed data into useful climate and geophysical information requires solving radiative transfer equations. Due to the need to process data at the scale of the entire planet, the methodologies currently used to approximate solutions of these complicated equations are limited and do not systematically exploit multi-sensor measurements, ground measurements, and the temporal dynamics at play. At the same time, countries around the world are increasingly turning to remote sensing data to cope with the challenges related to climate change and environmental degradation. To accurately inform stakeholders, better statistical models are needed, particularly for imaging of regions in the developing world where ground measurements are limited. The goal of this research is to develop a new generation of statistical methods for solving, at a more local level, the equations at the heart of satellite remote sensing data processing.The transformation of satellite-based remotely sensed data into useful climate and geophysical information requires solving some highly non-trivial radiative transfer inverse problems. This project aims to develop a Bayesian framework that combines algorithm unrolling deep learning models and a forward computer code into an inversion map. A transfer learning and a reinforcement learning framework will be developed to combine the inversion map learned in-silico with ground measurements, to adjust for distributional mismatch and to maintain the accuracy of the inversion procedure over time, even as the satellite data distribution changes over time. The project will also contribute at the theoretical level to a statistically deeper understanding of reinforcement learning and algorithm unrolling models. The project aims to improve analysis of global remote sensing data with applications to climate change, bridging of the disciplines of remote sensing, machine learning, and statistics.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.