基于机器学习的微波土壤水分产品降尺度研究

The microwave remote sensing is considered as one of the most effective ways to monitor soil moisture at large scales. However, the spatial resolution of currently available soil moisture products is on the order of tens of kilometers, which is not sufficient enough to meet the demand for various local-scale agricultural, meteorological, and hydrological applications. In order to solve such a problem, it is necessary to improve the spatial resolution of soil moisture products through downscaling methods. Statistical regression method is the most widely used downscaling method, which assumes that there is a simple regression relationship between soil moisture and land surface parameters. However, the relationship between soil moisture and surface parameters is nonlinear and highly complex, making the statistical regression method inaccurate. Machine learning methods can deal with complex non-linear relationships and show a great potential for soil moisture downscaling. This project mainly investigates the application of machine learning methods in soil moisture products downscaling. This research would build different machine learning models to determine the nonlinear relationship between soil moisture and surface parameters, and achieve downscaled soil moisture products. This project aims to develop a set of downscaling methods based on machine learning algorithms and to produce high resolution soil moisture products. The outcome of this research could provide high resolution soil moisture products for agricultural, meteorological and hydrological applications, and also helps to solve the problem of scale mismatch in validation of coarse resolution microwave soil moisture products, and consequently improve the monitoring capabilities in global climate changes and water cycle.

微波遥感是获取全球土壤水分最有效的手段之一。现有的土壤水分遥感产品的空间分辨率较低，不能满足农业、气象、水文等应用对于高空间分辨率土壤水分的要求。为了解决这一问题，需要对土壤水分进行降尺度以提高遥感产品的空间分辨率。统计回归方法是目前最常用的降尺度方法，假设土壤水分与地表参数之间存在简单的回归关系。然而，土壤水分与地表参数存在复杂的非线性关系，使得统计回归方法存在较大误差。机器学习可以模拟任何线性和非线性关系，在降尺度中具有巨大的潜力。本项目主要探讨机器学习方法在土壤水分降尺度中的应用，通过搭建不同的机器学习算法，拟合土壤水分和地表参数之间的非线性关系，实现降尺度。本项目将发展一套基于机器学习算法的土壤水分降尺度方法，用于高分辨率土壤水分产品的生产。本项目的研究成果不仅能满足不同应用对于土壤水分空间分辨率的需求，也有助于解决被动微波产品的验证难题，提高全球气候和水循环变化的卫星监测能力。

本项目针对现有土壤水分遥感产品空间分辨率较低，不能满足农业、气象、水文等应用对于高空间分辨率土壤水分的要求这一实际问题，开展了主被动微波融合土壤水分降尺度研究，以及基于机器学习的融合光学和微波数据的土壤水分降尺度方法研究。针对SMAP-Sentinel-1主被动微波数据的融合，评估了时间序列算法和物理模型算法，实现了高分辨率土壤水分的获取。针对光学和微波数据的融合，探讨了土壤水分与地表参数之间的关系，为搭建基于机器学习土壤水分降尺度奠定了基础。本项目的研究成果能够用于高分辨率土壤水分的获取。