Spatial Fusion of Himawari Observations and GFS Guidance for 24-72-Hour Regional Cloud-Probability Forecasting

We propose the Probabilistic Atmospheric Model (PAM) for regional cloud-probability forecasting at 24–72 h using Himawari observations and Global Forecast System (GFS) guidance. A recurrent convolutional encoder processes four six-hourly cloud observations, and a spatial fusion network blends its forecast with target-valid GFS guidance and applies a residual correction. The model uses approximately 0.59 million trainable parameters and predicts a 1° cloud-probability field over East Asia and the western Pacific. A post-hoc evaluation of previously viewed January–partial-July 2026 cases with three paired training seeds gives spatial Pearson correlations of 0.7587, 0.7111, and 0.6697 and pixel mean intersection over union (mIoU) of 67.45%, 64.13%, and 61.65% at 24, 48, and 72 h. Relative to scalar fusion, mIoU increases by 5.82–6.19 percentage points; area-weighted probability-field mean-squared error decreases by 22.64–24.45%. Field-score gains persist across six geographic sectors and groups defined by cloudiness and spatial variance. PAM also improves neighborhood agreement relative to raw GFS, while producing smoother fields and fewer recovered cloud objects. These results support competitive one- to three-day regional forecasting with compact satellite–numerical fusion. Comparisons with single-source, scalar, and single-component designs favor the complete spatial-fusion configuration.