FODNet: Fractional Orthogonal Disentanglement Network With Thick-to-Thin Suppression for Multilevel Cloud Detection

Accurate multilevel cloud detection is critical for a wide range of downstream applications in remote sensing imagery. However, spectral entanglement between clouds and complex backgrounds, together with the disparity in detection difficulty between thick and thin clouds, remains a key bottleneck limiting detection accuracy. To address these challenges, we propose a novel multilevel cloud detection framework based on fractional orthogonal decoupling and hierarchical thick-to-thin suppression strategies. Specifically, to disentangle cloud and background features, we introduce the fractional orthogonal disentanglement module (FODM). By incorporating a learnable fractional Fourier transform (Learnable FrFT), each FODM learns a layer-specific fractional order α , which is shared across all input images and optimized end-to-end over the training set, thereby identifying the most suitable observation domain for separating cloud features from background features at the corresponding feature hierarchy. Orthogonal subspace projection is subsequently employed to achieve effective separation. To mitigate the imbalance in detection difficulty between thick and thin clouds, we adopt a divide-and-conquer strategy by designing a dual-branch decoder that independently detects each type. A thick cloud response suppression module (TRSM) is introduced, wherein high-confidence predictions of thick clouds serve as reverse attention to explicitly suppress responses within the thin cloud branch. This mechanism compels the network to more effectively extract thin cloud signals from the remaining regions, thereby improving overall detection accuracy. Experiments conducted on L8-Biome and CloudSEN12 datasets demonstrate that the proposed method achieves superior detection performance compared with mainstream approaches, particularly in thin cloud regions.