DINO-Enhanced Feature Pyramid Network for Cloud Detection in Remote Sensing Images

Accurate satellite cloud image segmentation is a critical prerequisite for meteorological monitoring, radiation studies, and photovoltaic power station forecasting. However, traditional convolutional neural networks (CNNs) are limited by their local receptive fields, making it difficult to capture the global dependencies of large-scale cloud systems. Moreover, existing methods heavily rely on large-scale pixel-level annotations, which restricts their generalization capability under limited data. To address these issues, this letter proposes a satellite cloud image segmentation framework that integrates the self-supervised large model distillation with no labels (DINO) encoder with a feature pyramid network (FPN) decoder. The DINO model serves as the feature encoder, leveraging its superior feature extraction capabilities to capture cloud layer characteristics. The decoder combines FPN with U-Net skip connections, fusing multilevel features from the 2nd, 5th, 8th, and 11th layers to enable synergistic collaboration between shallow and deep information. Experimental results show that the proposed method achieves an intersection over union (IoU) of 93.98% on the HRC dataset and 95.07% on a self-constructed dataset, outperforming mainstream models such as UNet-ResNet50 and SegFormer.