Improving Winter Wheat Yield Estimation Under Saline Stress by Integrating Sentinel-2 and Soil Salt Content Using Random Forest

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Title: Improving Winter Wheat Yield Estimation Under Saline Stress by Integrating Sentinel-2 and Soil Salt Content Using Random Forest

This manuscript presents a promising study that integrates Sentinel-2-derived vegetation and salinity indices with field-measured soil salt content (SC) to estimate winter wheat yield using Random Forest (RF) models in Kenli County, China. The topic is relevant and timely, especially given the increasing challenges of salinity stress in agriculture. The idea of combining remote sensing indicators with ground-based data is well justified, and the results show strong model performance.

That said, the manuscript still needs significant revisions before it is ready for publication. Several methodological aspects are under-documented, and the interpretation of results could be improved. Below, I outline a few major and minor issues that I hope will help improve the clarity and impact of the work.

Major Comments:

1. The study does not account for basic soil physical properties such as clay, silt, and sand content. These can strongly influence the impact of salinity on crops. I believe this is a key limitation that should be clearly acknowledged in the discussion. The authors might also briefly explain how including these variables in the future could improve the robustness of the model.
2. While the modeling part is technically solid, it’s still not entirely clear how the results could be used in practice. Could this approach help with early-season yield forecasting or provide input to site-specific field management? A short paragraph explaining the potential applications for farmers or policymakers would be very useful.
3. The manuscript claims that salinity indices dominate during the P2 stage, but in Figure 5, NDVI seems more important. This contradiction needs to be addressed, either by clarifying the figure or adjusting the text.
4. The methods section would benefit from more detail about the RF configuration, e.g., mtry, ntree, nodesize, as well as how many predictors were used for each model variant (VI, SI, VI+SI, VI+SI+SC). A small table summarizing this information would be helpful for readers trying to replicate or understand the model structure.
5. The added value of including soil salt content (SC) is relatively small (e.g., R² improves from 0.77 to 0.78). Considering that SC is interpolated and computationally heavier, the authors should discuss whether this trade-off is worthwhile, especially when applying the model on a larger scale.
6. In Figure 9, the field sizes and spatial extent vary across subsets, which may make visual comparisons misleading. Please clarify the spatial extent in the figure caption and consider standardizing the comparison if that's the intended purpose.
7. The model is built and validated using data from a single season (2024). This limits its generalizability. It would be good to mention this as a limitation and suggest that future work include multiple years for more robust validation.
8. Just a note: feature importance in RF indicates how much a variable contributes to prediction accuracy, it doesn’t imply a causal effect. In a few places, the manuscript seems to treat importance scores as explanatory rather than predictive. I’d recommend softening the language here or clarifying this distinction.

Minor Comments:

1. The use of Ordinary Kriging for interpolating SC is fine, but the manuscript lacks details on semivariogram fitting and cross-validation. Including information on the variogram model (e.g., exponential, spherical), sampling density, and validation (e.g., RMSE) would improve transparency. Also, Kriging assumes stationarity, which may not apply in delta regions. You might want to briefly mention alternatives such as regression Kriging or co-Kriging.
2. Section 2.2.3 could be clearer, how many field samples were collected for classification? How were they distributed spatially? Including a confusion matrix in the supplementary material would also be helpful to evaluate the classification accuracy.
3. Several maps are missing scale bars or clear legends. Standardizing the color scales and adding map extents would make the visualizations easier to understand. A few brief figure notes could also help guide the reader.

Author Response

Please see the attachment. Thank you.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This is an interesting study on winter wheat yield estimation under saline stress using remote sensing data and the Random Forest algorithm. The paper is generally well-organized and clearly written. However, I have several questions and suggestions for the authors to consider:

1 Why was Random Forest selected as the primary modelling approach? While it is indeed capable of analysing feature importance, there are several other methods that offer similar capabilities. A justification for its selection would strengthen the methodology.

2. A comparative analysis is recommended to evaluate the performance of Random Forest against other commonly used methods in terms of both accuracy and computational efficiency.

3. The spatial distribution of winter wheat is derived from a separate Random Forest model. How is the uncertainty introduced by this upstream classification quantified and addressed in the yield estimation process?

4. Figure 9 requires a legend to facilitate interpretation and improve clarity.

Author Response

Please see the attachment. Thank you.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thanks for the corrections!

Reviewer 2 Report

Comments and Suggestions for Authors

The authors addressed all of my concerns and I have no more questions.