GACN: A Dual-Branch Graph Attention Convolutional Network for Cloud Detection Using Himawari-8/9 AHI

Clouds are core components of the global climate system and hydrological cycle, and high-precision cloud detection is a key prerequisite for quantitative optical remote sensing applications. Traditional threshold methods often miss thin clouds and misclassify high-albedo surfaces; existing pixel-level models ignore spatial context, causing class imbalance and detection bias. This study proposes the graph attention convolutional network (GACN) for daytime and nighttime cloud detection across land cover types using Advanced Himawari Imager (AHI) data. It adopts a dual-branch architecture to extract spatial correlations and multi-scale spectral features, paired with an adaptive fusion gate and multi-objective joint loss to balance detection performance. The graph attention branch builds dynamic spatial–spectral neighbor graphs to resolve spectral confusion between thin clouds and bright underlying surfaces. The proposed metric-constrained joint loss is also portable to other network architectures to mitigate class imbalance in pixel-level remote sensing classification. Trained on 2 km Himawari-8/9 AHI data with CALIPSO lidar labels, GACN achieves a TSS of 75.18% (daytime) and 71.57% (nighttime) with DR below 7%, outperforming seven baselines and two official AHI cloud mask variants. It provides an effective solution for high-accuracy AHI cloud mask generation and is more suited for regional-scale applications or offline post hoc scientific processing.