Cloud-top infrared observations reveal the four-dimensional precipitation structure

Referee: [Abstract] Abstract: The claim that 'cloud-top infrared measurements nevertheless encode sufficient information to recover the four-dimensional structure of precipitation' is supported solely by supervised performance on radar-derived targets. Because the network is trained to reproduce radar profiles, the reported 'recovery' constitutes a learned statistical mapping from co-located IR-radar pairs rather than an independent physical derivation from IR radiances alone. Without an ablation that removes radar supervision (e.g., training on IR only or on synthetic IR fields) or explicit out-of-distribution tests, the leap from 'predicts well on held-out radar' to 'physically encodes sub-cloud processes' remains under-supported.

Authors: We agree that the training procedure is supervised and that the reported performance reflects a learned mapping from paired IR-radar observations. At inference time, however, the network receives only multi-channel IR radiances and reconstructs the full 4D precipitation field; the ability to do so on held-out and independent radar datasets constitutes evidence that the requisite information is present in the IR observations. The moisture-first constraint is introduced precisely to reduce reliance on spurious training-set correlations. We will revise the abstract and add a dedicated discussion paragraph that explicitly states the supervised character of the training while emphasizing the inference-time use of IR alone and the supporting role of the physical constraint. We will also report additional out-of-distribution tests on data from different seasons and geographic regions. revision: partial

Referee: [Methods] Methods (moisture-first constraint description): The moisture-first constraint is presented as enforcing thermodynamic consistency by requiring latent recovery of precipitable water vapor. However, no ablation study quantifies its contribution to generalization, vertical structure accuracy, or performance outside the training distribution. Without such an ablation, it is unclear whether the constraint provides genuine physical anchoring or functions merely as an auxiliary reconstruction task that the high-capacity network can satisfy via training-set covariances.

Authors: We acknowledge that an explicit ablation would strengthen the justification for the moisture-first constraint. The constraint is motivated by the physical requirement that any realistic precipitation profile must be consistent with the column-integrated water vapor; the latent space is therefore regularized to reconstruct precipitable water vapor before predicting precipitation. In the revised manuscript we will add an ablation experiment that trains an otherwise identical network without the moisture constraint and reports the resulting changes in vertical profile accuracy, generalization error, and performance on independent radar data. revision: yes

Referee: [Results] Results (performance claims): The abstract states 'robust performance across large samples and independent radar comparisons' yet provides no quantitative error bars, RMSE or bias values, details on data exclusion criteria, or regime-specific breakdowns (e.g., light vs. heavy rain, tropical vs. mid-latitude). These omissions make it impossible to evaluate whether the reported skill is load-bearing for the 4D-structure claim or whether known radar biases (attenuation, sensitivity thresholds) propagate into the IR retrievals.

Authors: We apologize for the insufficient quantitative detail in the abstract and summary sections. The full results section contains performance statistics, but these were not summarized with error bars or regime breakdowns. In the revision we will (i) expand the abstract to include representative RMSE and bias values with uncertainty ranges, (ii) add a table that reports data exclusion criteria and performance stratified by rain rate and latitude band, and (iii) include a short discussion of how radar attenuation and sensitivity thresholds may affect the training targets and how the independent-radar validation helps assess propagation of those biases. revision: yes