Cloud detection via frequency-guided spatial modeling and bright surface suppression

Introduction Cloud detection in remote sensing imagery is critical for land cover mapping, time-series analysis, and environmental monitoring. Existing methods often struggle to detect optically thin clouds, preserve fine boundaries, and avoid false positives over bright non-cloud surfaces such as snow, rooftops, and deserts. Methods We propose a novel framework integrating frequency-domain information with spatial modeling. A 2D Fast Fourier Transform (FFT) extracts frequency-aware features through a Frequency-guided Spatial Attention Module (SFAM), enabling better capture of high-frequency details and boundary structures across cloud types including thick cumulus and semi-transparent cirrus. A Bright Surface Confusion Reducer (BSCR) dynamically suppresses false activations over highly reflective backgrounds by exploiting distinct spectral and textural characteristics of snow, urban surfaces, and bare soil. An edge-body cooperative supervision strategy further refines segmentation through consistency-aware boundary optimization. Results Experiments on three benchmark datasets spanning multiple satellite platforms— Landsat-8 Biome, MODIS, and Sentinel-2—demonstrate MIoU improvements of 3.12%, 1.34%, and 4.24% respectively over state-of-the-art baselines, with substantial false positive rate reductions across all surface types including snow/ice (15.1% absolute reduction). Discussion The proposed architecture offers a practical solution for high-precision cloud detection across diverse remote sensing scenarios, generalizing effectively across sensors with spatial resolutions ranging from 10m to 1km without requiring sensor-specific tuning.