CloudDeepLabV3+: a lightweight ground-based cloud segmentation method based on multi-scale feature aggregation and multi-level attention feature enhancement

The segmentation of ground-based cloud images is the basis for obtaining numerous parameters. To achieve high-precision adaptive image requirements, this study designs a lightweight method named CloudDeepLabV3+ that integrates multi-scale features aggregation and multi-level attention feature enhancement. Firstly, novel EfficientNetV2-S designed as extraction backbone to reduce network Secondly, heterogeneous receptive field splicing atrous spatial pyramid pooling module designed. It enhances correlation information between layers, realizes multiscale fusion. enhancement based on self-attention mechanism intensifies representation local global features. Thirdly, alignment constructed pull deep shallow alignment. Finally, we implement ablation key components comparison experiment with other advanced methods using five evaluation metrics. Results show play an important role in promotes accuracy while reducing loss detailed Generalization performance verification indicates excellent proposed model cloud-mask generation.