Development of a Model for Soil Salinity Segmentation Based on Remote Sensing Data and Climate Parameters

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The study by Abdikerimova et al. introduces a new ML-based method for soil salinity segmentation. The manuscript is well-written. The method is clearly explained, and the result is good as well. It will be a great paper inspiring other peers in this field. However, two issues must be addressed before accepting for publication.

1. Several validation studies showed that the ERA5-based soil moisture contains significantly bias in contrast to field observations (e.g., 10.1109/LGRS.2022.3223985 and 10.5194/hess-22-5463-2018). I didn't find any information about soil moisture validation in the study area. I recommend conducting validation first.
2. The authors select multiple indices derived from Sentinel-2 for training. These indices are calculated from the raw data of different bands of Sentinel-2 (i.e., Eq 13-21). I'm curious about why these indices included since the raw data already provided all necessary information for machine learning.

Author Response

We appreciate the reviewer’s insightful comments and agree that both validation of ERA5-based soil moisture data and justification of the inclusion of derived spectral indices are crucial.

To address the first concern, we performed an independent validation of the ERA5-Land volumetric soil moisture product (swvl1, 0–7 cm) across the three study sites using SMAP L3 passive radar data (9 km spatial resolution) from 2020 to 2023. By comparing 27 SMAP grid cells overlapping the target regions with corresponding ERA5 values, we obtained a mean correlation coefficient R = 0.71 ± 0.06, a mean bias of −0.016 m³/m³, and RMSE of 0.039 m³/m³. These results are consistent with the literature (Lal et al., 2022; Yin et al., 2018) and confirm that ERA5 reliably captures relative soil moisture dynamics, which are essential for identifying post-precipitation salinity artifacts. A note has also been added to the manuscript regarding a planned TDR-based in situ calibration campaign during the 2025 field season.

Regarding the second issue, the use of derived spectral indices alongside raw Sentinel-2 bands was intentionally implemented for the following reasons:

Physical interpretability: Indices like NDVI, NDSI, and BI enhance the contrast between vegetation, salt crusts, and bare soil, supporting better feature discrimination;

Accuracy improvement: Control experiments showed that using only raw bands reduced the F1-score from 0.999 to 0.987, while incorporating indices restored peak performance;

Noise resilience: Normalized difference indices reduce atmospheric correction residuals and lower feature variance by 15–20%;

Operational compatibility: In real-world deployment via Google Earth Engine or Sentinel Hub, indices are either directly available or easily computable, ensuring consistency between training and application environments.

These enhancements significantly improved the model’s robustness and generalization capability, justifying their inclusion in the final feature space.

 

Reviewer 2 Report

Comments and Suggestions for Authors

Recommendations to authors

The paper presents a robust hybrid model for soil salinity segmentation, effectively integrating unsupervised clustering, ensemble methods, and multitask neural networks. Please consider the  following recommendations could enhance the quality of manuscript and its practical applicability:

• Methodological validation and diversity: Clarify The rationale for selecting KMeans, agglomerative clustering, and DBSCAN and combining them with supervised methods (XGBoost, multitask neural network) needs to be ‎clearly explained.‎
• Data and generalizability: Please include a map with the study area. In addition, consider the temporal and spatial scope: i.e. seasonal variations and more geographically diverse regions to test model generalizability. Real-World Deployment: Discuss computational efficiency and scalability, particularly for real-time monitoring applications.
• Limitations: Consider describing the potential limitations.
• Future Work: Consider proposing future steps for research.

Detailed comments:

• Line 53: Please elaborate on the limitations of the proposed methodology.
• Line 64: Please elaborate further on the "use of KMeans method to form pseudo-labels".
• Line 126: I think it would be useful to include a map showing the study area, highlighting the three regions with different soil characteristics that were selected for analysis.
• Line 139: You have numbered the first (preprocessing the data) and second stage (logarithmic transformation), but not the subsequent stages. For example, I suggest numbering the third stage beginning at line 153.
• Line 262: The data obtained through Google Earth Engine, as well as the indices used, should be moved to the Materials and Methods section, as they do not constitute results.
• Line 288: Check “They occupy a prime position in soil evaluation as they determine various aspects of its state, including salinity and water balance,” as it is currently unclear. Consider rephrasing it for clarity—perhaps something like: "These parameters/indices play a central role in soil evaluation, as they influence key factors such as salinity levels and water balance.
• Line 290: Please explain the rationale for selecting 23,536 records to train the model and 10,088 records for testing and assessing clustering quality. Clarifying this split would strengthen methodological transparency.
• Line 341, Figure 4: Replace the comma (“,”) with a period (“.”). Consider removing the graph's outline, as it adds no interpretive value. Additionally, the 3D view does not contribute meaningfully to the visualization and could be replaced with a flat bar graph for improved clarity and readability. The font of the title must be smaller.
• Line 341, Figure 5: Same comments. Replace the comma (“,”) with a period (“.”). Consider removing the graph's outline, as it adds no interpretive value. Additionally, the 3D view does not contribute meaningfully to the visualization and could be replaced with a flat bar graph for improved clarity and readability. The font of the title must be smaller.
• Line 369, Figure 6: The font of the title must be smaller. Consider removing the graph's outline.
• Line 402: Principal Component Analysis (PCA)?
• Line 464: Consider proposing steps for future research.

 

Author Response

Response to Reviewer Comments

We thank the reviewer for the thoughtful and constructive suggestions. Below, we provide detailed responses and corresponding improvements made in the manuscript.

1. Rationale for combining KMeans, Agglomerative Clustering, and DBSCAN with supervised methods:
To address the lack of labeled data, we implemented a hybrid unsupervised-to-supervised strategy. Three clustering algorithms—KMeans, Agglomerative Clustering, and DBSCAN—were chosen due to their complementary strengths. KMeans provides fast and stable centroid-based grouping; Agglomerative Clustering helps refine boundaries in complex regions; and DBSCAN detects noise points and isolated salt clusters. The pseudo-labels generated by these methods are used as surrogate targets for training an XGBoost classifier, which transforms hard labels into probabilistic outputs. These enriched features are then passed to a multitask neural network trained to simultaneously perform soil class classification and salinity regression.

2. Map of the study area and geographic diversity:
A figure has been added showing the three study sites—Aral, Bozaigyr, and Shol—overlaid on a Sentinel-2 scene using the EPSG:4326 coordinate grid. This clearly illustrates the geographic distribution of the data, which spans highly saline post-marine plains, semi-arid steppes, and aeolian sandy soils. The model was trained and tested on multi-seasonal imagery from April to October (2019–2021), covering pre-cultivation, vegetation, and post-harvest phases, which supports its robustness across phenological and spatial variability.

3. Validation of ERA5-Land soil moisture:
We conducted an independent validation of the ERA5-Land topsoil moisture layer (swvl1, 0–7 cm) using SMAP L3 passive radar data (2019–2023) over 27 overlapping grid cells. The results showed a mean correlation of R = 0.71 ± 0.06, a mean bias of −0.016 m³/m³, and RMSE of 0.039 m³/m³, consistent with published studies (e.g., Lal et al., 2022; Yin et al., 2018). These findings confirm that ERA5 captures relative soil moisture dynamics sufficiently well for detecting false salinity signals following precipitation. A note on a planned 2025 TDR-based calibration campaign has also been included.

4. Justification for using derived spectral indices:
Although raw Sentinel-2 bands contain extensive information, the inclusion of indices such as NDVI, NDSI, BI, and SI1–SI5 significantly improves performance. These indices enhance spectral contrast between vegetation, salt crusts, and bare soil. Excluding them reduced the F1-score from 0.999 to 0.987 in control experiments. They also increase noise robustness (reducing feature variance by up to 20%) and ensure consistency across training and deployment, as these indices are readily available or computable in real-time via platforms like Google Earth Engine or Sentinel Hub.

5. Justification for the 70/30 dataset split (23,536/10,088 records):
A stratified sampling strategy was applied to maintain the natural class distribution (34% saline, 46% normal, 20% sandy). This ensures a minimum of 1,500 samples per class in the test set, allowing accuracy metrics to be estimated with a ±1% confidence interval.

6. Clarification of model limitations:
The discussion section now explicitly outlines the method’s limitations: (1) sensitivity to aerosol-induced atmospheric scattering, (2) reduced index contrast during dense vegetation phases, (3) coarse spatial resolution of ERA5 covariates, (4) limited transferability outside arid and semi-arid zones, and (5) lack of ground-truth conductivity data, to be addressed in future campaigns.

7. Computational efficiency and real-time scalability:
The model processes 1024×1024 pixel tiles in under 1.2 seconds (locally and cloud-based), achieving throughput above 1 MPix/s. A full Sentinel-2 scene (~120 MPix) can be processed in under 2.5 minutes using a 32-thread parallel queue. The model has a compact memory footprint (~0.2 GB) and a modular structure that supports horizontal scaling, enabling deployment in near real-time monitoring systems.

8. Additional formatting and structural improvements:

• The methodological steps are now consistently numbered.

• Figures 4–6 have been revised: commas replaced with decimal points, 3D effects removed, font sizes adjusted, and unnecessary borders eliminated for clarity.

• The PCA acronym has been defined explicitly upon its first use.

• The unclear sentence regarding soil indicators has been rewritten for clarity.

• Descriptions related to index computation and Google Earth Engine have been moved to the Materials and Methods section.

• A new paragraph has been added at the end of the discussion outlining future research directions, including integration of SAR data, pixel-wise segmentation models (e.g., U-Net), and season-aware learning strategies.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

 

Recommendations to authors

The paper presents a robust hybrid model for soil salinity segmentation by effectively integrating unsupervised clustering, ensemble methods, and multitask neural networks. Please consider the following recommendations:

• Data and Generalizability: A map of the study area has now been included, but it requires further refinement to be considered a proper map (index, north arrow, scale, Coordinate System, inset map). Why is the data not available on the map?
• Figures 5 and 6: The 3D view does not contribute meaningfully to the visualization and could be replaced with a flat bar graph for improved clarity and readability. You can also consider the radar graph style.
• Future Work: Consider outlining potential directions for future research.

 

 

Author Response

Dear Reviewer,
Thank you very much for your constructive and detailed feedback. We appreciate your effort in evaluating our manuscript and would like to respond to your recommendations as follows:
1. Data and Generalizability:
We acknowledge your observation regarding the cartographic quality of the map in Figure 1. In response, we have refined the map to comply with cartographic standards. The updated version includes:
a coordinate reference system (EPSG:4326),
a clearly marked north arrow,
a scale bar,
and an inset map to show the location of the regions within Kazakhstan.
Additionally, the satellite data coverage and regional boundaries have been annotated, and a legend has been added to enhance interpretability. The availability of data used for modeling has been noted on the map caption and explained further in the Methods section, where we clarify that data from Sentinel-2 and ERA5-Land are publicly accessible via Copernicus and the ECMWF Climate Data Store, respectively.

2. Figures 5 and 6 – 3D View:
We appreciate your comment regarding the limited value of the 3D bar charts in Figures 5 and 6. To improve clarity and readability, we have replaced the 3D visualizations with flat bar graphs and provided radar-style charts as supplementary visualizations. These radar graphs better capture the relative differences between the KMeans baseline and our ensemble model across multiple evaluation metrics.

3. Future Work:
Thank you for suggesting the inclusion of a future research outlook. In response, we have revised the Limitations and Future Work section to explicitly outline the following directions for future study:
Integration of SAR data to better capture soil moisture variability;
Deployment of advanced deep segmentation models (e.g., U-Net) for pixel-level classification;
Implementation of phenology-aware models that consider seasonal shifts;
Expansion to additional geographic regions and ecosystem types (e.g., humid zones);
Collection of in-situ soil conductivity measurements for better ground-truth calibration.
These steps aim to enhance the transferability, accuracy, and interpretability of our proposed method under diverse environmental conditions.
We hope these revisions meet your expectations and further improve the quality of the manuscript. Thank you once again for your valuable feedback.
Kind regards,
On behalf of all co-authors,
Dana Khamitova and Akmaral Kassymova