Continual Learning for Geostationary Satellite Remote Sensing of Precipitation

Abstract

Accurate precipitation retrieval from geostationary satellites remains a formidable challenge, primarily due to sensor limitations and the complexity of mapping satellite observations to surface precipitation. To address these challenges, this research proposes a deep learning (DL)–based framework for precipitation estimation using GOES-R satellite measurements. However, precipitation characteristics can vary substantially across geographic regions, storm types, and climatic conditions, causing DL models trained in one domain to perform poorly in another. Therefore, continual learning (CL) is needed to enable the model to learn from new precipitation domains while preserving knowledge acquired from previous domains. In this work, new CL algorithms have been developed to improve model generalization across multiple precipitation retrieval tasks. Three different incremental algorithms are introduced to efficiently update the learned representation as new domains are incorporated, enabling scalable adaptation without retraining from scratch. Building upon this framework, we design a CL-based precipitation retrieval system that explicitly accounts for multi-domain variability by allowing a single model to sequentially learn from different study domains while maintaining performance on previously learned domains. Experimental results demonstrate that: (1) the proposed DL-based framework enhances current operational precipitation products derived from GOES-R satellites; (2) the proposed CL algorithm improves the generalization capability of DL models, allowing them to perform robustly across multiple tasks; and (3) the integrated CL-based precipitation framework further strengthens performance across diverse study domains spanning the southeastern United States. These findings highlight the potential of the proposed approach as an important step toward scalable DL-based precipitation estimation systems for broader geostationary satellite applications.