Abstract The recent and projected investments across the world on radar satellite missions (e.g., Sentinel-1, SAOCOM, BIOMASS and NISAR) provide a great opportunity for operational radar soil moisture mapping with high spatial and temporal resolution. However, there is no retrieval algorithm that can make complementary use of the multi-frequency data from those missions, due to the large uncertainties in observations collected by the different sensors, different validity regions of the forward models, and the fact that inversion algorithms have been designed for specific data sources. In this study, the principle of ensemble learning was introduced to provide two general soil moisture retrieval frameworks accounting for these issues. Instead of trying to find an optimal global solution, multiple soil moisture retrievals (termed sub-retrievals) with moderate performance were first obtained using different channels and/or time instances randomly selected from the available data, with the retrieved ensemble of results being the final output. The ensemble retrievals, taking one existing snapshot method and two multi-temporal methods as the base retrieval algorithms, were evaluated using a synthetic data set with the effectiveness confirmed under various uncertainty sources. An evaluation using the Fifth Soil Moisture Active Passive Experiment (SMAPEx-5) data set showed that the ensemble retrieval outperformed the non-ensemble retrieval in most cases, with a decrease of 0.004 to 0.014 m3/m3 in Root Mean Square Error (RMSE) and an increase of 0.01 to 0.16 in correlation coefficient (R). Weakly biased and correlated sub-retrievals were confirmed to be the basic requirement of an effective ensemble retrieval, being consistent with use of ensemble learning in other applications.

中文翻译：

---

用于多 SAR 任务土壤水分反演的随机集成方法

摘要 世界各地最近和预计对雷达卫星任务（例如 Sentinel-1、SAOCOM、BIOMASS 和 NISAR）的投资为具有高空间和时间分辨率的业务雷达土壤水分测绘提供了绝佳机会。然而，由于不同传感器采集的观测值存在较大不确定性，正演模型的有效区域不同，以及反演算法对这些任务的多频数据进行互补，目前尚无反演算法可以互补利用这些任务的多频数据。专为特定数据源而设计。在这项研究中，引入了集成学习的原理，以提供两个解决这些问题的一般土壤水分反演框架。而不是试图找到一个最优的全局解决方案，首先使用从可用数据中随机选择的不同通道和/或时间实例获得具有中等性能的多个土壤水分反演（称为子检索），检索到的结果集合是最终输出。集合检索以一种现有的快照方法和两种多时态方法作为基础检索算法，使用合成数据集进行评估，并在各种不确定性来源下验证了有效性。使用第五土壤水分主动被动实验 (SMAPEx-5) 数据集进行的评估表明，在大多数情况下，集合检索优于非集合检索，均方根误差 (RMSE) 降低了 0.004 至 0.014 m3/m3相关系数 (R) 增加 0.01 至 0.16。