Increasing the spatial coverage and temporal resolution of Earth surface monitoring can significantly improve forecasting or monitoring capabilities in the context of smart city, such as extreme weather forecasting, ecosystem monitoring and anthropogenic impact monitoring. As an essential data source for Earth’s surface monitoring, most satellite observations exist data gaps due to various factors like the limitations of measuring equipment, the interferences of environments, and the delay or loss of data updates. Although many efforts have been conducted to fill the gaps in the last decade, the existing techniques cannot efficiently address the problem. In this paper, we extensively study the gap-filling problem of satellite observations using imbalanced learning. Specifically, we propose a framework called Reanalysis to Satellite (R2S) to simulate satellite observations with reanalysis data. In the R2S framework, we propose a generic method called Spatial Temporal Match (STM), matching reanalysis data and satellite observations to construct the Reanalysis-Satellite (R-S) dataset used to train the model. Based on the R-S dataset, we propose a novel method called Semi-imbalanced (SIMBA) to handle the imbalance problem of gap-filling by taking advantages of traditional machine learning and imbalanced learning. We construct a hybrid model in the R2S framework for the Soil Moisture Active Passive (SMAP) satellite observations of the tropical cyclone wind speed. Extensive experiments demonstrate the hybrid model outperforms the traditional machine learning model and closely approximates in situ observations.

增加地球表面监测的空间覆盖范围和时间分辨率可以显着提高智慧城市环境下的预测或监测能力，例如极端天气预报，生态系统监测和人为影响监测。作为地球表面监测的重要数据源，由于各种因素（例如测量设备的局限性，环境的干扰以及数据更新的延迟或丢失），大多数卫星观测都存在数据缺口。尽管在过去的十年中已经进行了许多努力来填补空白，但是现有技术不能有效地解决该问题。在本文中，我们使用不平衡学习广泛研究了卫星观测的空白问题。特别，我们提出了一个名为“卫星再分析”（R2S）的框架，以利用再分析数据模拟卫星观测。在R2S框架中，我们提出了一种称为空间时空匹配（STM）的通用方法，该方法可以对重新分析数据和卫星观测值进行匹配，以构建用于训练模型的重新分析卫星（RS）数据集。基于RS数据集，我们提出了一种称为半不平衡（SIMBA）的新方法，以利用传统机器学习和不平衡学习的优势来处理缺口填充的不平衡问题。我们在R2S框架中为热带气旋风速的土壤水分主动被动（SMAP）卫星观测构建了一个混合模型。广泛的实验表明，混合模型优于传统的机器学习模型，并且非常接近原位观察。