This paper maps rice paddy distribution of East Asia countries in 2023 at 10m resolution (EARice10), it is a timely and accurate (according to the assessment) product that can help estimate greenhouse gas emissions and grain yield. The paper developed a new method that integrates the use of both Sentinel-2 Optical and Sentinel-1 SAR data. The paper is in general well-written with high-quality figures.

I have a few comments that I hope can help improve the manuscript.

My major concern is about the independent validation sample (91320). The authors briefly mentioned they are obtained through field surveys and visual interpretations.  How much is obtained through field surveys? How visual interpretation can accurately obtain validation samples? It needs details about visual interpretations and shows that is robust.  I doubt how can one visually distinguish rice paddies from other crops.

Eq (5), it needs more details that how coefficient B and A are determined, based on what comparison and criteria. Are these empirically derived values B and A needs change when applied to a different region, for example South Asia? If so, how others adopt this method can determine the values.

Line 203, Page 11: “ divide the plots in to different objects”, this sentence is very confusing. What are the plots referring to? Is there an image segmentation done or not? Do optical images and SAR images both go through the same procedure of image segmentation? It not sure if the final classification is based on pixels or a group of pixels- objects.

Line 205, Page 11, why size parameter of 15, compactness of 0.8, and connectivity of 8. It seems the SAR images are segmented; how do you treat the optical images then?

Line 236-237, page 13. What are these 2000 samples used for? Is it for training the following random forest model? Are they verified?  I can see these 2000 samples have a very high probability of being rice if both Rice_Optical and Rice_SAR=1, but still no 100% guarantee.

3.2 Optical-SAR adaptive fusion model. This part, while it is creative, it is not necessary and seems too complicated to me. So about 24 RF models are trained for semi-month mean features of Sentinel-2 images. What are the input image features, are these the same as Table 1? What are the training labels (from 3.1.3?), are all RF models trained using the same set of training labels?

When a model is trained on each semi-month mean image S-2, you lose the temporal dynamic information as shown in Figure 6. I am not sure of the reasoning behind doing this, or whether it has any advantages. Second, why would the model bother running on cloudy pixels, why not just masking cloudy pixels out?

How is the model trained to combine P_Optical, N_cloud-free, and P_SAR? Again, training labels.

Line 346, Page 20. This refers to my previous comments, more details are needed to convince the independent samples are reliable.

Summary of Comments:

This paper maps rice paddy distribution of East Asia countries in 2023 at 10m resolution (EARice10), a timely and accurate product (according to the assessment) that is useful for estimating greenhouse gas emissions and grain yield. The method integrates both Sentinel-2 Optical and Sentinel-1 SAR data. The manuscript is generally well-written, with high-quality figures. However, I have a few concerns that I hope will help improve the manuscript:

Major Comments:

Validation of Data:

The independent validation samples (91,320) are said to be obtained through field surveys and visual interpretations. How many of these samples were obtained through field surveys?

Visual interpretation as a method for distinguishing rice paddies from other crops lacks detail and robustness. More information should be provided to justify that visual interpretation can accurately identify rice paddies, or field survey photos should be presented as evidence.

Application to Fragmented Rice Fields (Southern China):

Given that the method uses an optical-SAR adaptive fusion model, it may struggle to perform well in fragmented rice fields, especially in mountainous southern China. The small, scattered fields and complex terrain in this region could lead to errors in classification.

It is recommended to incorporate additional geographical data (e.g., terrain, soil types) to limit the model and improve its performance in such areas. The model's adaptability to these factors could significantly enhance its accuracy in more complex regions.

Comparison with Existing Data:

While the manuscript emphasizes the improved resolution of the rice map (10m), there are no comparisons made with prior datasets, particularly those from 2015-2021, which were at 500m resolution.

The authors should consider comparing the distribution and precision of the current data with these earlier datasets to demonstrate the advancements achieved through EARice10. This would help showcase the superior results of this study in a more concrete manner.

Specific Suggestions:

Equation 5 – Normalization of SAR Features:

In Equation (5), where the SAR features are normalized, the manuscript does not clearly explain how the preset parameters A and B were derived. Are these values empirical, and if so, how were they selected? The paper should provide more justification or sensitivity analysis showing how different parameter values affect the results, particularly in regions like southern China where terrain and cropping patterns differ from other areas.

Additionally, if these parameters were tuned based on a specific region, would they need to be regionally adjusted when applying the method to other areas (e.g., South or Southeast Asia)? Clarifying this would improve the reproducibility of the approach.

SRMI Threshold:

The SRMI threshold for rice classification (set at 0.5) may not be flexible enough to handle the variability in terrain and crop conditions, particularly in fragmented and mountainous regions. It would be useful to conduct a sensitivity analysis on this threshold to see how changes in SRMI values affect classification accuracy in different landscapes. This analysis could highlight the robustness of the SRMI index across diverse geographic regions.

Random Forest Model Training:

In Section 3.2, where multiple Random Forest models are trained for each semi-monthly image, there is a risk of over-complicating the model and losing temporal dynamics, as indicated by Figure 6. The paper should clarify why it was necessary to train multiple models rather than aggregating data over a longer period to maintain temporal information. This section could also benefit from a discussion on overfitting risks, particularly in fragmented regions where training data may be more heterogeneous.

It is recommended to test a more streamlined approach that reduces the number of models and retains more temporal information to potentially improve classification performance, especially in areas with sparse data coverage.

Validation and Field Survey Data:

The validation data is largely based on visual interpretation. However, the manuscript lacks details on how this interpretation was conducted, especially in regions with fragmented rice fields. Providing more concrete evidence, such as field survey data or ground truth photographs, would strengthen the validation process. The authors should also clarify the proportion of samples that were obtained through actual field surveys versus visual interpretation.

Furthermore, it would be useful to provide accuracy metrics for different regions, especially comparing flat and mountainous areas, to demonstrate the model’s robustness across diverse landscapes.

Publisher’s note: this comment is a copy of RC2 and its content was therefore removed on 11 October 2024.

Publisher’s note: this comment is a copy of RC2 and its content was therefore removed on 11 October 2024.