基于不适定反问题求解的降水图像降尺度研究

Study on Precipitation Image Downscaling Based on the Method of Ill-posed Problems Solving

遥感降水产品和环流模型预报降水降尺度研究一直是水文和气象学的热点。使用多源降水融合资料进行降水图像降尺度研究,其本质是提高低分辨率观测或模拟降水场分辨率,并适当增加其细节或高频特性。基于降水自相似结构性质,将不适定数学反问题求解法用于降尺度。在求解降尺度不适定反问题时,不同风暴环境的小规模组织内降水特征往往会重复出现这一性质,通过训练得到高、低分辨率的降水场,形成相应的"完备字典",用于正交匹配追踪贪婪法重构高分辨率降尺度的降水场。执行时,首先基于专家场(Fields of Experts,FoEs)模型进行缺测资料插补;其次采用文中方法进行降水图像降尺度应用研究。基于传统"保真度"度量指标和空间结构相似性度量法对该方法得到的试验结果进行评估,结果表明该方法可行。

Downscaling of remote sensing precipitation products and the forecasting of circulation model are al- ways the intense interests in hydrology and meteorology. The essence of downscaling is primarily to enhance resolu- tion of observation or simulated rainfall field, and to appropriately increase its details or high frequency characteristics. Precipitation, as the main driving factors of the earth＇ s hydrologic cycle, not only affects the moisture and heat condition of a certain river basin, but also affects the global water and heat circulation. Based on the properties of rainfall self-similarity structure, the mathematically ill-posed precipitation problem solving method was used in low resolution downscaling precipitation for high resolution reconstruction. When solving the downscaling ill-posed problem, the greedy method of orthogonal matching pursuit was introduced so as to get the best high-resolution esti- mation in an optimal sense. It is hard to imagine that we might be able to find very similar （in mathematical norms） precipitation patterns over relatively large storm-scales. However, finding similar features over sufficiently small sub-storm scales seems more feasible. Based on the characteristics that small scale organized precipitation fea- tures tend to recur across different storm environments, the precipitation of both high and low resolution was ob- tained by training, which could be used to reconstruct the desired high-resolution precipitation field. Multi-source merged precipitation products were used in this experiment. Given the consideration of incompleteness of merged precipitation dataset, it was firstly interpolated based on the method of Fields of Experts （FoEs） , which could solve the problem that common interpolation method could hardly work on the interpolation for dataset where consecutive missing data exists. Secondly, ideal experiments of precipitation products downscaling were carried out, where smooth coupling sampling and resampling operator were adopted respectively. Assessment based on the metrics of fidelity and spatial structural similarity demonstrates that the method used in this paper is feasible.