A stand-alone framework for predicting spatiotemporal errors in satellite-based soil moisture using tree-based models and deep neural networks

Soil moisture (SM) has been recognized as one of the essential climate variables in Earth science. Incorporating spatiotemporal information from satellite SM observations into land surface models via land data assimilation (LDA) has emerged as a promising method to improve continuous SM modeling and climate extreme monitoring. Since the reliable satellite SM error dynamics are crucial for successful LDA applications but often assumed to remain static over time, this study presents a novel framework that combines triple collocation analysis (TCA) with advanced machine learning techniques, such as Light Gradient Boosting Machine (LGBM) and Deep Neural Networks (DNN), to accurately quantify spatially and temporally continuous satellite-based SM characteristics on a global scale. In this study, the stand-alone TCA-based time-variant SM error prediction models, which rely exclusively on data used for the Soil Moisture Active Passive (SMAP) retrieval, were developed and comprehensively evaluated. These models successfully recover error information over areas where SMAP SM error data were previously unusable due to uncertain error characteristics. In addition, the stand-alone SMAP-based model demonstrates superior performance compared to model that relies on external datasets, such as the Global Land Data Assimilation System (GLDAS). These findings provide valuable insights into the dynamic nature of satellite-based SM error under various environmental conditions and present a novel way to improve LDA. Furthermore, the proposed methodology can be extended to predict error dynamics for other satellite-based geophysical datasets, broadening its potential applications beyond SMAP.土壤湿度（Soil Moisture, SM）已被公认为地球科学领域的关键气候变量之一。通过陆面数据同化（Land Data Assimilation, LDA）将卫星土壤湿度观测的时空信息融入陆面模型，已成为改善连续土壤湿度模拟与气候极端事件监测的极具前景的方法。可靠的卫星土壤湿度误差动态对成功应用陆面数据同化至关重要，但过往研究通常假设其随时间保持静态。为此，本研究提出一种全新框架，将三重collocation分析（Triple Collocation Analysis, TCA）与轻量梯度提升机（Light Gradient Boosting Machine, LGBM）、深度神经网络（Deep Neural Networks, DNN）等先进机器学习技术相结合，以精准量化全球尺度下基于卫星的时空连续土壤湿度特征。本研究中，研究人员仅依托土壤湿度主动被动（Soil Moisture Active Passive, SMAP）反演所用数据，开发了纯基于三重collocation分析的时变土壤湿度误差预测模型，并开展了全面评估。这些模型成功恢复了此前因误差特性不明而无法获取SMAP土壤湿度误差数据的区域的误差信息。此外，相较于依托全球陆面数据同化系统（Global Land Data Assimilation System, GLDAS）等外部数据集的模型，纯基于SMAP的模型展现出更优的性能。本研究结果为理解不同环境条件下基于卫星的土壤湿度误差的动态特性提供了宝贵见解，并为改进陆面数据同化提供了全新路径。进一步而言，所提出的方法学可拓展至预测其他基于卫星的地球物理数据集的误差动态，从而拓宽了其应用场景，不限于SMAP。