MT-HCCAR: Multi-Task Deep Learning with Hierarchical Classification and Attention-Based Regression for Cloud Property Retrieval

Abstract

In Earth science, accurate retrieval of cloud properties including cloud masking, cloud phase classification, and cloud optical thickness (COT) prediction is essential in atmospheric and environmental studies. Conventional methods rely on distinct models for each sensor due to their unique spectral characteristics. Recently, machine/deep learning has been embraced to extract features from satellite datasets, yet existing approaches lack architectures capturing hierarchical relationships among tasks. Additionally, given the spectral diversity among sensors, developing models with robust generalization capabilities remains challenging for related research. There is also a notable absence of methods evaluated across different satellite sensors. In response, we propose MT-HCCAR, an end-to-end deep learning model employing multi-task learning. MT-HCCAR simultaneously handles cloud masking, cloud phase retrieval (classification tasks), and COT prediction (a regression task). It integrates a hierarchical classification network (HC) and a classification-assisted attention-based regression network (CAR), enhancing precision and robustness in cloud labeling and COT prediction. Experimental evaluations, including comparisons with baseline methods and ablation studies, demonstrate that MT-HCCAR achieves optimal performance across various evaluation metrics and satellite datasets.