This study introduces a novel Mix-n-Scale framework with GNNs for short-term air temperature (Ta) forecasting at the intra-city scale. The authors convincingly demonstrate that incorporating spatial interactions improves forecast skill, achieving more than a 12 % reduction in RMSE compared with a conventional LSTM-only baseline. They provide a thorough evaluation of model performance and present well-structured analyses. Overall, this is a solid paper and is a meaningful contribution with clear potential to advance short-term forecasting in urban areas. However, there are several areas where additional experiment and discussion would strengthen the manuscript. I would support publication after minor revisions.

Major comments:

1. The authors show that adding GNN-based spatial aggregation improves performance, but it is unclear how much of the gain depends on the number, proximity, and connectivity of stations. I suggest including the following analyses to provide deeper insight into the value of spatial information for Ta forecasting and to offer practical guidance for applying the framework in settings with sparser or differently configured station networks.
   • Conduct a sensitivity analysis by removing stations one by one and evaluate how forecast performance changes. For example, I would like to know the effectiveness of inclusion of station 2 on station 3 given their distance. This would also help justify the assumption that “clear directed relationships for information propagation from specific ‘super-nodes’ may not exist” (Lines 236–237).
   • Verify whether the observed performance gains arise from the GNN’s relational structure rather than simply having access to neighbor data. Consider comparing with a simpler baseline where neighbor embeddings are concatenated or averaged and fed directly to the final prediction layer without using GNN.

1. Discuss why including global predictors appears to worsen GSAGE performance (Fig. S1) and explain the rationale for keeping them.
2. Since the manuscript discusses drawbacks of reanalysis-trained DL models (Lines 49 – 51), please add a brief benchmarking context (e.g., a small literature table or paragraph) showing how your error metrics compare with these approaches.
3. The ensemble approach is a key novelty and clearly improves performance over any single model. It would be helpful to know how sensitive the results are to the ensemble size or to variations in the hyperparameters, as well as how computational resources scale with ensemble complexity to inform an optimized training strategy. What are the best-performing topologies of the GNN and how different are they from each other?
4. The explanation of the spatiotemporal pattern of forecast performance in Section 3.4 could be clearer. While the authors report RMSEs, local Ta variability, and site characteristics, they do not clearly connect these characteristics to the observed performance differences. For instance, the statement that “the pronounced diurnal contrast can be primarily attributed to solar radiation-induced perturbations and consequent atmosphere-land interactions, highlighting the inherent challenges in capturing daily peak values” (Lines 323 – 325) notes the difficulty but does not explain why these processes make the peak harder to forecast. I recommend revising this section to make those linkages more explicit.

Minor comments:

1. Latitude and longitude swapped in Table 1.
2. Since GAT and GSAGE themselves are GNNs without LSTM encoding, the labels or figure captions could be clarified to avoid misunderstanding. For example, Fig. 1 may read as though you are comparing a standalone GNN with an LSTM, rather than GNNs applied on top of LSTM embeddings.
3. Line 211: revise “the mean Ta patterns is …” to “the mean Ta patterns are …”.
4. Line 266: revise “significantly outperform than …” to “significantly outperform the …”.

This work demonstrated that graph neural networks (GNNs) can harnessobservation network information to refine Ta prediction at individual locations and elucidate underlying mechanisms. The topic is very interesting and has important implications in prediction of urban weather. The paper is well organized and written. The findings of this study are worth of publication in the ACP after minor revision as following:

1. The rolesof Urban morphology, local circulation, anthropogenic heat release and land use and their impacts on local climate in the Hong Kong should be compared with other regions, that is, can LOI defined in the present work be related to these local background factors? Just for wind or altitude?

Effects of anthropogenic heat release upon the urban climate in a Japanese megacity; Diurnal variation of amplified canopy urban heat island in Beijing megacity during heat wave periods: Roles of mountain-valley circulation and urban morphology; Simulating the Regional Impacts of Urbanization and Anthropogenic Heat Release on Climate across China;Unevenly spatiotemporal distribution of urban excess warming in coastal Shanghai megacity, China: Roles of geophysical environment, ventilation and sea breezes;Adjustment of the urbanization bias in surface air temperature series based on urban spatial morphologies and using machine learning; Impacts of Various Local Climate Zones on Canopy Urban Heat Island and Dry Island: Spatial Heterogeneity and Relative Contributions in Beijing

2. About overfitting issues of machine learninglearning models, how did you address or aovid it?

3. Station ID can be removed in Figures 2 and 6b for clear show.