Efficient Retrieval of Subpixel Cloud Fraction From Coarse Cloud Masks Using Dual‐Branch Deep Learning

Abstract Cloud detection is a critical procedure in satellite remote sensing. Most meteorological satellite products provide binary cloud masks, which identify whether a pixel in the satellite image is entirely cloudy or clear‐sky, and the subpixel cloud fraction (CF) of partly cloudy pixels is usually unavailable. In this work, we develop a deep‐learning neural network to estimate the subpixel cloud fraction of moderate‐resolution satellite images based on binary cloud masks, and train the neural network with high‐resolution satellite data. We aggregate 10 m Sentinel‐2 cloud masks into physically consistent 1 km binary cloud masks, smooth them with edge‐aware filtering to obtain collocated 1 km CF reference fields, and use these pairs to train DualCloudNet, a dual‐branch U‐Net that fuses complementary feature representations. The network takes the aggregated 1 km cloud mask as input and predicts the corresponding 1 km continuous CF, achieving a Pearson correlation of 0.92 and an RMSE of 0.103 on an independent test data set. The model outperforms a physically motivated statistical estimator, and its accuracy is slightly higher than two other deep‐learning networks that were constructed in the validation stage of this study. Applied without retraining to FY‐3D MERSI‐II and Landsat‐8 imagery, DualCloudNet reproduces pixel‐level cloud structures, demonstrating cross‐sensor generalization. Compared with existing cloud fraction retrieval algorithms, the new algorithm is able to provide a subpixel cloud fraction with low computational cost. The fine‐scale CF information is potentially useful for improving cloud property retrievals and surface radiation budget estimation, where the representation of partly cloudy pixels remains a major source of uncertainty.