Continual Lifelong Learning for Precipitation Retrievals Using ABI and GLM Measurements on the GOES-R Series

Accurate precipitation retrieval using satellite sensors remains challenging due to limitations in spatiotemporal sampling of the satellite sensors and the uncertainty associated with the applied parametric retrieval algorithms. In response to this challenge, deep learning (DL) frameworks are designed for precipitation retrieval using satellite observations. However, traditional DL algorithms can only learn from the data in the pre-defined study domain and may not be adaptive to the new data from new domains, whereas these domains may have different terrain characteristics. To address this issue, we propose a continual lifelong deep learning framework for continual precipitation retrieval over two distinct domains using the observations on the geostationary operational environmental satellite (GOES) series. For the DL model, the cloud-top brightness temperature from multiple advanced baseline imager (ABI) channels and the lightning flash rate from geostationary lightning mapper (GLM) measurement are used as inputs to the deep learning model, and the precipitation products from multi-radar multi-sensor (MRMS) system are used as target labels to optimize the deep learning model. For developing a continual lifelong learning framework, we first train a baseline DL model over a pre-defined study domain, then the elastic weight consolidation (EWC) technique is used to adjust the DL model parameters when we shift the previously trained DL model to a new domain. This EWC approach allows the DL model to be adaptive to the data over the new domain while not causing any catastrophic forgetting of the precipitation patterns in the previous domain. The performance of the DL models is evaluated across different rainfall intensities for light, moderate, and heavy precipitation scenarios. The experimental results show that the adjusted DL model using EWC can achieve a compromising performance over both domains. These findings suggest that the proposed continual lifelong learning framework for the DL model can effectively offer accurate precipitation retrieval over various domains.