| 19 Aug 2025
AGCPP: All-day Global Cloud Physical Properties dataset with 0.07° resolution retrieved from geostationary satellite imagers covering the period from 2000 to 2022
Abstract. The use of remote sensing to accurately measure cloud properties and their spatial and temporal variability has become an important area of atmospheric science research. However, the heterogeneity of data formats across national agencies and the calibrate and navigate associated with the use of data from different agencies have prevented the climate research community from using the full continuum of global cloud physical properties products. In this paper, All-day Global Cloud Physical Properties (AGCPP) is proposed, which provides cloud physical properties covering nearly the entire globe, from latitude -70° to 70° and longitude -180° to 180°. The main attributes of this dataset include cloud phase, cloud top height, cloud optical thickness, and cloud effective radius, with a time range from 1 January 2000 to 31 December 2022. AGCPP combines the observational advantages of geostationary satellites and polar-orbiting satellites. It uses the Moderate Resolution Imaging Spectroradiometer (MODIS) Level-2 cloud product (MOD06/MYD06) to train the cloud-based attention-UNet (CloudAtUNet) model, and then evaluates AGCPP using MOD06/MYD06 and the Cloud–Aerosol Lidar with Orthogonal Polarisation (CALIOP) 1 km cloud layer product. The evaluation results indicate that AGCPP demonstrates excellent continuity and consistency in both temporal and spatial accuracy, as well as high consistency in diurnal accuracy. Due to the long time series and all-day global nature of the dataset, it is expected that the dataset AGCPP will significantly increase the potential for climate change research, particularly with respect to potential feedback effects between clouds, surface albedo, and radiation. AGCPP is stored in the Network Common Data Format (netCDF), a standard that allows various tools and libraries to process the data quickly and easily. The AGCPP dataset is freely available on the Science Data Bank at https://doi.org/10.57760/sciencedb.26292 (Zhao et al., 2025), and the corresponding code can be found at https://github.com/lingxiao-zhao/AGCPP (last access: 25 June 2025).
This study develops an AI algorithm for cloud-top phase, altitude, and particle size, and cloud optical thickness based on the infrared window band (~11 μm) and infrared water vapor band (~6.7 μm) from geostationary satellite observations, alongside temperature & humidity profile products from reanalysis data. The research topic is interesting. However, all of the input variables of the deep learning model are solely correlated with cloud-top height, the inversion of other cloud properties lacking physically explainability (see below for details). Numerous technical details remain underspecified. At minimum, a major revision is required before consideration for acceptance.
Major Comments:
1. Physical explainability:
Fundamentally, neither the ~11 μm infrared window band nor the numerical ERA-5 temperature -humidity profiles provide direct information about cloud-top phase, particle size, or cloud optical thickness—they primarily constrain cloud-top height. While incorporating the ~6.7 μm infrared water vapor band may marginally improve cirrus cloud-top height retrieval, these parameters inherently lack sensitivity to phase/microphysics. Consequently, the inversion model lacks robustness. For instance, although cloud-top temperatures (brightness temperatures) above 0°C typically indicate liquid phase and below -40°C suggest ice phase, values between -40°C and 0°C can represent supercooled liquid, ice, or mixed-phase conditions. The proposed empirical relationships, trained on large datasets, constitute merely statistical correlations—analogous to predicting human height from tree height—with tenuous physical connections. This limitation explains the significant performance drop observed when validating the phase classification algorithm independently on 2022 data. Further independent testing (exposing unseen scenarios) would likely yield even lower metrics. Authors must establish stronger physical justification for their methodology to convince reviewers and readers.
2. Data matching
Considering the disparate spatial resolutions—8 km for geostationary satellite products, 0.25° for reanalysis data, versus 1 km for polar-orbiting MODIS/CALIOP—how are these multi-source datasets temporally and spatially matched? Has horizontal Homogeneity been addressed?
How does the vertical cloud-phase profiling capability of CALIOP map onto the single-layer cloud-top phase retrieved from geostationary imagery? Given CALIOP’s inability to penetrate optically thick clouds, is its reported optical thickness suitable as a ‘truth’ reference for evaluating your algorithm’s performance?
Minor Comments:
1. When first introducing “cloud effective radius” in the abstract and main text, use the precise term“ cloud-top particle effective radius” or “cloud-top particle size” .
2. Spell out the full names upon first occurrence for acronyms like CARE, CLARA.
3. 2.1.2: Specify the exact MODIS product ID(s) and version(s) used. Note that MODIS offers three types of cloud-top particle effective radius at 1.6 μm, 2.1 μm, and 3.9 μm, please clarify which one was selected.
4. Table 4 & Fig. 4: Compared to prior studies, the reported accuracies are relatively low. Particularly, the coefficient of determination (R²) for cloud optical thickness is notably poor.