Fine‐Tuning a Machine‐Learned 3D Cloud Climatology Reveals Aspects of Cloud Cover Trends

Abstract The Chalmers Cloud Ice Climatology (CCIC) is unique among long‐term cloud records: using retrievals from merged geostationary 11 μm observations, it provides continuous 3D estimates of both frozen hydrometeor mass and cloud probability. Here, we present a key update to CCIC: enhanced detection of thin clouds through tuning of its neural network cloud probability outputs. The update delivers robust information from local instantaneous retrievals to long‐term, large‐scale averages. We compare multidecadal, height‐resolved trends in cloud cover from CCIC and ERA5, revealing subtle but emerging long‐term changes in total cloud cover. By exploiting CCIC's capabilities, the contributions of different cloud types and thicknesses to these changes become clearer. This helps reconcile inconsistent trends between purely observational data and ERA5, suggesting that discrepancies stem from varying sensitivities to high, thin clouds.