Hybrid Spatial Analysis of Rurban Dynamics Using Geospatial and Socio-Economic Data: Case of Casablanca–Settat Region

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Recommendations to the Authors

Some points requiring improvement:

• Check the format of all the tables. The format of tables in the manuscript is inconsistent with the standards of the journal.
• Check all the references used in the text; some are provided in text format and does not align with the manuscript's numerical bracket system.
• Regarding figures: all images need further improvements. More specific: (1) in figure 6, please remove the pie graphs. Also increase font size; (2) in figure 10 please increase image size or font size.

Detailed comments:

• Line 37: creates new genetically altered rurban forms" [1,3] Morocco – Missing punctuation. A period is required before "Morocco".
• Line 54: between these categories.Existing; missing space after the period.
• Line 69: 2.1. Defining urban areas – Heading capitalization is inconsistent with title case conventions used in subsequent sections.
• Line 85: The exponential decay equation requires sequential numbering to the right margin per standard journal guidelines.
• Line 140: éparpillée [1976]; Citation format deviates from the manuscript's numerical bracket system.
• Line 197: (MLC) [Richards, 1986]; Citation format is inconsistent with the numerical reference standard.
• Line 191: decision trees [Quinlan, 1986]; Citation format is inconsistent with the numerical reference standard.
• Lines 294-295 Regarding 0 = Urban; 1 = Peri-urban; 2 = Rurban; 3 = Rural. Use a sentence instead of a structured vertical list.
• Lines 653-654, Figure 9. Classification Confidence Levels of Communes 2014;2024. Obviously, the caption is wrong. The text label for Figure 9 duplicates the terminology of Figure 6 instead of providing a unique caption that accurately characterizes the Potential Agricultural Score distribution discussed in the accompanying paragraphs.

Additional References:

Chen, Y., Zhao, P., Lin, Y., Sun, Y., Chen, R., Yu, L., & Liu, Y. (2024). Semantic-Enhanced Graph Convolutional Neural Networks for Multi-Scale Urban Functional-Feature Identification Based on Human Mobility. ISPRS International Journal of Geo-Information, 13(1), 27. https://doi.org/10.3390/ijgi13010027 ‎

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors This manuscript addresses an important and timely topic—the delineation of rural-urban continuum dynamics in the Casablanca–Settat region—by integrating Sentinel-2 imagery, census data, and machine learning techniques. The hybrid framework combining Self-Organizing Maps (SOM) and Graph Neural Networks (GNN) with expert-based decision rules is conceptually appealing, and the ten-year diachronic perspective (2014–2024) offers valuable insights into the territorial transformations underway in this rapidly urbanizing Moroccan region. Overall, the study is well-motivated and the cartographic outputs are visually compelling, yet several methodological and presentational issues need to be addressed before the paper can reach its full potential. My primary concern lies in the justification and performance of the GNN component. The manuscript reports a GNN-to-rules agreement of merely 53% and a Cohen’s Kappa of approximately 0.31, which falls into the "fair" agreement category at best. While the authors acknowledge that GNNs capture spatial dependencies and neighborhood interactions, the marginal concordance with expert classifications raises serious questions about the practical utility of this model within the proposed hybrid architecture. The reader is left wondering whether the GNN adds genuine value or merely introduces unnecessary complexity. I would urge the authors to either provide a much stronger theoretical or empirical justification for retaining the GNN—perhaps by demonstrating that it captures specific transitional patterns missed by SOM—or to reconceptualize its role, for example, as an auxiliary sensitivity analysis rather than a coequal classification branch. Additionally, the sharp degradation in SOM performance between 2014 (Kappa = 0.842) and 2024 (Kappa = 0.428) is noted but insufficiently explained. Is this decline attributable to genuine territorial blurring, increased spectral heterogeneity, or instability in the expert rules themselves? A deeper diagnostic investigation, possibly including sensitivity tests on input variables, would strengthen the argument considerably. The feature extraction and multi-source fusion strategy also warrants closer scrutiny. The current pipeline normalizes twelve socio-environmental indicators and reduces dimensionality via PCA before feeding them into SOM and GNN. While this is a conventional approach, the manuscript does not adequately discuss whether the linear assumptions underlying PCA are appropriate for the highly non-linear, heterogeneous patterns characteristic of peri-urban and rurban landscapes. More importantly, recent advances in geospatial representation learning suggest that adaptive, non-linear feature reconfiguration can substantially improve the integration of multi-resolution and multi-modal inputs. For example, intelligent learning reconfiguration models based on optimized transformer architectures and multi-source feature supervision have demonstrated considerable advantages in preserving spatial detail and modeling complex feature interactions from heterogeneous remote sensing data, as shown in recent work on high-precision InSAR DEM reconstruction (https://doi.org/10.1109/TGRS.2025.3535546). Although the application domain differs, the underlying principles of adaptive multi-source fusion and learned feature reconfiguration are directly relevant to the challenges encountered here, particularly regarding the integration of spectral indices with coarse-resolution census variables. I encourage the authors to discuss how such advanced feature learning paradigms might address the spectral-socioeconomic fusion problem in future iterations of their framework. From a structural standpoint, Section 4.6 on agricultural suitability assessment feels somewhat disconnected from the main narrative thread. The preceding sections establish a coherent argument about territorial classification and urbanization dynamics, but the introduction of the Potential Agricultural Score (PAS) and AHP-based weighting appears abruptly and lacks clear linkage to the classification results. If the authors wish to retain this section, they should clarify how the PAS analysis informs, or is informed by, the four-class urbanity typology. For instance, do peri-urban communes with high PAS values face greater development pressure? Conversely, are rural-to-rurban transitions correlated with declining agricultural potential? Without such explicit connections, this section reads like a separate study and dilutes the manuscript’s focus. Either integrate it more tightly or consider removing it to maintain thematic coherence. The reliance on commune-level aggregation, which the authors briefly acknowledge as a limitation in the conclusions, deserves more critical attention throughout the paper. Given that peri-urban and rurban areas are frequently characterized by fine-scale spatial heterogeneity—where built-up parcels, agricultural plots, and vacant land coexist within a single administrative unit—aggregating all indicators to the commune scale inevitably masks intra-communal variation. This spatial smoothing may explain why transitional zones exhibit low classification confidence and why the SOM’s separability metrics deteriorate in 2024. The authors should discuss this scale issue explicitly in the discussion section and consider whether sub-communal sampling or dasymetric mapping techniques might mitigate the problem, even if only as a recommendation for future research. Finally, the transparency and reproducibility of the expert-based rule set need improvement. While Figure 1 provides a schematic overview, the actual thresholds, weighting coefficients, and hierarchical logic of the decision tree are described only in general terms. For a study that places considerable emphasis on the hybrid fusion of data-driven and expert knowledge, the lack of a fully specified rule matrix—or at minimum, a supplementary table detailing the percentile-based thresholds for each indicator and class—makes independent verification difficult. I strongly recommend that the authors provide these details, either in the main text or as supplementary material, to ensure that the methodology can be replicated in other contexts. In summary, this paper tackles a genuinely interesting problem and offers a rich, multi-source perspective on Moroccan urbanization dynamics. However, the uneven performance of the GNN, the underdeveloped linkage between classification and agricultural suitability, the scale limitations of commune-level analysis, and the insufficient detail on expert rules currently hold the work back. Addressing these issues would not only strengthen the manuscript’s scientific rigor but also enhance its practical utility for planners and policymakers. I look forward to seeing a revised version.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors This manuscript addresses the detection and classification of the rural‑urban continuum in a rapidly urbanizing Moroccan region. The integration of Sentinel‑2 spectral indices, GHSL population layers, census socio‑demographic data, and two complementary AI models (SOM and GNN) is genuinely innovative. With several targeted improvements, the paper will make a valuable contribution to the fields of remote sensing, urban geography, and territorial governance.b Specific Comments as follows:   1 The introduction effectively describes the Moroccan context, but it could benefit from a stronger opening statement on the global urgency of understanding rurbanization and peri‑urbanization (e.g., UN SDG 11, land degradation, food security). This would help position the study within the international literature. 2. Sections 2.1–2.4 provide a solid historical and conceptual background. However, Section 2.5 (“Contribution of Deep Learning and Machine Learning”) is somewhat long and includes detailed descriptions of SOM and GNN that partly repeat the methods section. Consider condensing this section and moving technical details to the methods. 3. Section 3.6 (“Graph Neural Network for classification”) appears twice (page 10) – the second occurrence seems to be a duplicate heading with incomplete content. Please merge or remove. 4. The GNN description lacks specific details on the number of GCN layers, activation functions, learning rate, and validation strategy. Please add these for reproducibility. The SOM grid sizes differ between 2014 (8×8) and 2024 (6×5) – why? This should be justified. The expert rule‑based thresholds (Section 3.4) are derived from percentiles – please clarify whether the same percentiles were used for both years or recalibrated. 5. The validation results (Tables 2 and 3) are informative but raise questions. For SOM, the Silhouette score drops from 0.142 to –0.036 in 2024. Negative values indicate that many points may be assigned to the wrong cluster – does this mean the 2024 classification is unreliable? The authors argue that this reflects “increasing overlap due to urban continuity”, which is plausible, but they should add a stronger caution in the discussion. Also, the Cohen’s Kappa for SOM drops from 0.842 (almost perfect) to 0.428 (moderate) – this is a dramatic change. Please discuss whether this is due to data heterogeneity or a limitation of the SOM algorithm for more continuous gradients. 6. Section 4.4 concludes that SOM is more suitable than GNN because of better interpretability and expert agreement. This is a reasonable choice given the study objectives. However, the GNN achieved only ~53% agreement with expert rules. The authors could acknowledge that GNN might be improved with more training data or a different graph construction (e.g., using distance‑weighted edges instead of binary contiguity). This would be a fair and constructive comment. 7. The extracted thresholds for each class (NDVI, population, illiteracy, etc.) are very useful for planning. However, note that for the Rurban class (Table 7), the schooling rate is shown as “90.0–90.5%” – a very narrow range that seems unlikely given the standard deviation. Please check the original data; this might be a rounding or calculation error. If correct, explain why the range is so tight. 8.The integration of agricultural potential mapping is a nice addition, but it feels somewhat disconnected from the main classification. Please clarify how the PAS (Potential Agricultural Score) relates to the four territorial classes. For example, do high‑PAS areas correspond exclusively to rural communes? The text mentions that “rural communes demonstrate high PAS indices” – a figure showing PAS distribution by class would be helpful.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Thank you very much for your research. However, I would like to recommend several improvements that should be addressed in order to strengthen the international scientific representativeness and overall quality of your study.

• All equations should include proper numbering and a clear description of the parameters used.
• In the Introduction or Section 2.3, there is no analysis of the evolution of peri-urban areas in different countries worldwide. Including an international perspective would help to identify current global trends, compare differences among countries or continents, and better contextualize your study within the broader scientific discussion.
• The literature review is currently too descriptive and does not provide a strong scientific positioning of your study within the global body of knowledge related to this problem. While the terminology is adequately described, it is also necessary to present and critically discuss previous quantitative findings from related studies.
• All figures should include a brief description of their content within the main text.
• More detailed information about the datasets used in the research should be provided, including, at minimum, the acquisition dates of the satellite images.
• The maps are missing scale bars.
• Figure 7 has very poor quality, which makes it difficult to properly evaluate its content.
• The discussion section should be significantly strengthened. You should include the limitations of the research and compare your findings with similar studies conducted in other countries and regions of the world. Additionally, it would be valuable to discuss whether your methodology and dataset could be applied in other contexts, such as Latin America or Europe.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report

Comments and Suggestions for Authors

Thank you for your modifications to the document. In its current state, I can now recommend it for publication.