This study combines multi-source remote sensing data with residual neural networks (ResNets) to develop the first 30 m resolution annual China Forest AGB dataset (1985–2023) with uncertainty quantification. It’s a long-term series of AGB data, which  is of great significance for estimating forest carbon sink in China.

Comments:

​1. Introduction​

• ​The first three paragraghs need to be condensed.
• Line 90-100, It is best to compare the advantages and disadvantages of the existing data. eg: Have you compared international time series data, such as the AGB data of the European Space Agency, etc.
• Why did you choose the ResNet-based deep learning algorithm? You should make it clear, because this is your core technical approach.

​2. Methodology & Data​

• ​The GEDI AGBD samples are between 2019 to 2021, and field survey data across China is from the 1990s to the 2010s. How did you deal with the gap in time.
• Given that GEDI data has only been available since 2019, how to ensure the model's predictive ability for the early years (1985-2018) requires a more detailed explanation.

​3. Results ​

• ​Spatial pattern of forest AGB: Figure 6 shows the spatial pattern of AGB in 2023, but it lacks corresponding diagrams for earlier years (such as 1985), making it difficult to visually compare the changes. It is suggested to supplement the AGB distribution map of key years.
• Trends in AGBD/AGC: Critique the reported increase in AGBD (95.74 to 122.69 Mg/ha) and AGC stock (5.50 to 13.97 PgC). Are these trends plausible given China’s afforestation efforts?
• ​Drivers of Change: Evaluate the partitioning of AGC growth (65.1% forest growth vs. 34.9% expansion). Are these proportions supported by ancillary data (e.g., land-use maps)?

​4. Discussion​

• ​You compared trends of different datasets, your data has a long time series, is there more accurate for every year? I think you’d better use field data to verify.
• Although the comparison with GEDI data showed a smaller deviation (0.06 Mg/ha), the comparison with ground survey data showed a larger negative deviation (-19.87 Mg/ha), especially in the tropical rainforest area (-68.57 Mg/ha). This systematic deviation requires more in-depth analysis and discussion.
• Although the paper discusses the influence of different canopy coverage thresholds on AGB estimation in lines 547-555, it does not quantitatively explain the degree of influence of using different thresholds (such as 10% vs. 20%) on the main conclusions of this study. It is suggested to add sensitivity analysis.

​

This manuscript generated a long-term 30 m Annual Aboveground Biomass data from 1985 to 2023 in China. They mainly used a residual neural networks (ResNets) together with the GEDI footprint AGBD and Landsat images to generate the AGBD map. The manuscript is generally well-written. I have several suggestions listed below, which I think may be helpful for revising the manuscript.

Materials and methods:

1. Does the observations from Landsat suffers from the sensor degradation effect? Since they are merged by different sensors, it may have some impacts on the long-term trend of Landsat data, which can finally propagate to the AGBD map.
2. How do you consider the risk of overfitting of your model? And the multicollinearity issue among the predictors you choose?
3. How do you validate the long-term trend of AGBD you generated? Since it is the key novelty aspect of this study.

Results and discussion:

1. Figure 3: from this figure, it suggests that your AGBD map is systematically lower than the observed value, right? Since the slope is lower than one, and with a negative bias value. Please explain this point.
2. Figure 8: how do you validate that the temporal trend in AGBD you generated is correct or not? Could you find some ground observations to validate this?
3. Figure 9: how do you separate the AGBD trend into forest expansion and forest growth? Please added some detailed process of how you did this.
4. Figure 15: we can find a large difference (even two-three times larger) between your map with Hengeveld, CCI and Chen’s maps. Could you explain the underlying reason?