Review article: Hydrologically Enhanced Machine Learning Framework for Urban Flood Inundation Mapping Using Multi-Sensor Remote Sensing Data: A Case Study of Mumbai, India

Pawar, Ankush S.; Phade, Gayatri M.

doi:10.5194/egusphere-2026-1275

Preprints

https://doi.org/10.5194/egusphere-2026-1275

Preprints

22 Apr 2026

| 22 Apr 2026

Status: this preprint is open for discussion and under review for Natural Hazards and Earth System Sciences (NHESS).

Review article: Hydrologically Enhanced Machine Learning Framework for Urban Flood Inundation Mapping Using Multi-Sensor Remote Sensing Data: A Case Study of Mumbai, India

Ankush S. Pawar and Gayatri M. Phade

Abstract. The complicated terrain, highly populated building surfaces and insufficient credible ground observations make urban flood mapping difficult in urbanizing megacities that rapidly develop in coastal areas. This study suggests that a hydrologically improved machine learning architecture can be utilized to perform automated urban flood inundation mapping by combining multi-sensor satellite data with a scalable decision support system (DSS). The Google Earth engine used Sentinel-1 SAR, Sentinel-2 optical imagery, SRTM digital elevation data, and CHIRPS precipitation data to create a comprehensive predictor stack.

To explicitly model flood propagation controls that most data-driven models tend to omit, two new hydrologic-topographic predictors were created:-the Relative Elevation Model (REM) and River Network Index (RNI), to model local terrain depressions and hydraulic connectivity. A consensus-based combination of SAR backscatter change, optical water indices, and topographic constraints produced flood labels with approximately 2.6x10⁵ pixels of floods in the Mumbai Metropolitan Region during the 2019 monsoon season. A representative training set was formed using balanced stratified sampling for use in the supervised classification. Random Forest, optimized XGBoost and ensemble models were created and tested in Python using official classification measures. The tuned XGBoost model had the best performance with an overall accuracy of 71.7 percent and an area under the receiver operating characteristic curve (AUC) of 0.803, which performed better than the Random Forest and ensemble configurations. The statistical significance of the improvement in model discrimination was at the 95 percent confidence level. The analysis of ablation revealed that the model discrimination of REM and RNI increased by approximately 5–6 percent in AUC, which proves their importance in urban flood detection. There is high spatial congruency between the predicted inundation pattern and known flood-prone regions along the major drainage patterns.

The proposed framework provides a reproducible, scalable, and hydrologically informed framework for urban flood inundation mapping and has high potential for operational flood monitoring and decision support in data limited tropical cities.

Received: 07 Mar 2026 – Discussion started: 22 Apr 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Ankush S. Pawar and Gayatri M. Phade

Status: open (until 31 Jul 2026)

Post a comment Subscribe to comment alert

CC1:
'Comment on egusphere-2026-1275', Anupal Baruah, 26 Apr 2026 reply

1) I am somewhat skeptical about the use of SAR data for urban flood mapping, given the well-known double-bounce scattering effect in built-up areas, which can lead to misclassification of flooded regions. This limitation is already acknowledged in Table 1. In light of this, it would be helpful if the authors could further justify their decision to proceed with SAR data, and clarify how they mitigate or account for these uncertainties in their analysis.
2) Additionally, the use of 30 m spatial resolution may be too coarse for accurately capturing urban flood dynamics, where fine-scale features such as roads, drainage networks, and building footprints play a critical role

Reply

Citation: https://doi.org/10.5194/egusphere-2026-1275-CC1
- AC1:
  'Reply on CC1', Gayatri M Phade, 26 Apr 2026 reply
  We sincerely thank the commenter for the insightful and constructive feedback on our manuscript. The points raised regarding the use of SAR data in urban environments and the implications of spatial resolution are highly relevant, and we appreciate the opportunity to clarify these aspects.
  
  Regarding the use of SAR data in urban flood mapping, we agree that double-bounce scattering in built-up areas can introduce uncertainties and may lead to misclassification of flooded regions. Despite this limitation, SAR data remain a widely adopted and valuable source for flood mapping due to their all-weather, day-and-night imaging capability, which is particularly critical during flood events characterized by cloud cover.
  
  In our study, we mitigate these limitations through a multi-source data integration framework. Specifically, SAR-derived features are not used in isolation but are combined with hydrologic–topographic indicators such as relative elevation and flow accumulation, as well as additional geospatial predictors. This integration allows the machine learning model to reduce reliance on any single data source and improves robustness against SAR-specific artefacts. Furthermore, the use of statistical descriptors and temporal variability features helps to distinguish true flood signals from urban backscatter effects. We will further clarify this mitigation strategy in the revised manuscript.
  
  With respect to the use of 30 m spatial resolution, we acknowledge that this resolution may not fully capture fine-scale urban features such as roads, drainage networks, and individual buildings. However, the choice of 30 m resolution was guided by the need to maintain consistency across multiple datasets (e.g., DEM, precipitation, and derived hydrologic variables) and to ensure computational feasibility for regional-scale analysis.
  
  Our objective is to provide a scalable and generalizable framework for urban flood susceptibility mapping rather than detailed street-level inundation modelling. The machine learning framework leverages terrain and hydrologic context, which remain meaningful at this spatial scale. Nevertheless, we agree that higher-resolution datasets could further enhance the accuracy of flood delineation in dense urban environments. This limitation and its implications will be more explicitly discussed in the revised manuscript, along with suggestions for future work using higher-resolution data.
  We thank the commenter again for these valuable suggestions, which will help improve the clarity and robustness of our study.
  
  Reply
  
  Citation: https://doi.org/10.5194/egusphere-2026-1275-AC1
RC1:
'Comment on egusphere-2026-1275', Anonymous Referee #1, 31 May 2026 reply

The manuscript addresses an important and timely topic in urban flood inundation mapping by integrating multi-sensor remote sensing data with machine learning and hydrologic-topographic predictors. The use of Sentinel-1 SAR, Sentinel-2 optical imagery, SRTM DEM, CHIRPS rainfall, and HydroRIVERS data provides a relevant basis for developing a scalable flood mapping framework. The inclusion of the Relative Elevation Model (REM) and River Network Index (RNI) is a promising attempt to improve the physical interpretability of machine learning-based flood detection. The reported performance of the tuned XGBoost model, with an accuracy of 71.7% and AUC of 0.803, suggests that the proposed framework has potential for operational urban flood monitoring. However, the manuscript requires substantial revision before it can be considered scientifically robust. First, the title describes the paper as a “Review article,” but the manuscript is clearly an original research article involving data processing, model development, model evaluation, and case study application. This should be corrected to avoid confusion.
The novelty of the study also needs to be clarified. The manuscript claims that REM and RNI are new hydrologic-topographic predictors, but it does not sufficiently explain how these indices differ from existing flood conditioning variables such as elevation, slope, distance to river, HAND, drainage proximity, topographic wetness index, or flow accumulation. The authors should clearly state whether REM and RNI are newly developed indices, modified versions of existing indices, or case-specific hydrologic features. A major concern is the formulation of the RNI. The text describes RNI as a measure of hydraulic proximity and connectivity to river networks, but the equation uses cumulative precipitation divided by the elevation difference from the minimum DEM. This formulation does not directly represent distance to river, drainage connectivity, or river network influence. The authors should revise the RNI equation so that it is mathematically consistent with its stated hydrological meaning.
The flood label generation process also requires stronger justification. The manuscript uses a consensus rule based on SAR backscatter ratio, backscatter difference, NDWI, and REM thresholds, but the selected threshold values are not adequately justified. The authors should explain why SAR ratio > 1.25, backscatter difference ≥ 3 dB, NDWI > 0.05, and REM < 5 m were selected. A threshold sensitivity analysis would strengthen the reliability of the generated flood labels. The validation strategy is another important limitation. Since the training and testing labels are generated from remote sensing-based consensus rules, the model may be learning the labeling assumptions rather than being validated against independent flood observations. The authors are encouraged to include independent validation data, such as official flood records, observed flood locations, high-resolution imagery, or historical flood-prone areas in Mumbai. If such data are unavailable, the manuscript should clearly state that the reported accuracy reflects agreement with consensus-generated labels rather than confirmed ground truth.
The statistical significance claims should also be improved. The manuscript states that the XGBoost model significantly outperforms other models based on the DeLong test, but the p-values, confidence intervals, and test statistics are not reported. Since the AUC difference between Random Forest and XGBoost is small, these values are necessary to support the claim of statistical significance. There is also an inconsistency in the reported ensemble performance. Table 3 reports the RF-XGB ensemble AUC as 0.794, while the ROC figure appears to show a different ensemble AUC value. The authors should carefully check and correct all performance values in the tables, figures, and discussion.
Finally, the manuscript requires substantial language editing. Several sentences are awkward or unclear, and the citation style is inconsistent between author-year and numbered formats. The figures, especially the workflow and spatial flood maps, should be improved for readability and publication quality. Figure 5 should include clearer legends, units, class definitions, and map elements. Overall, the study has potential, but the current version needs major revision. The authors should strengthen the novelty statement, correct the RNI formulation, justify the flood-label thresholds, improve validation, report full statistical testing results, resolve inconsistencies in model performance, and substantially revise the language and presentation.

Reply

Citation: https://doi.org/10.5194/egusphere-2026-1275-RC1
- AC2: 'Reply on RC1', Gayatri M Phade, 01 Jun 2026 reply
  
  We thank Referee #1 for the detailed and constructive review of our manuscript. We appreciate the positive assessment of the overall framework and the valuable suggestions for improvement.
  We acknowledge the concerns regarding manuscript classification, the novelty and formulation of the hydrologic–topographic predictors, threshold selection, validation strategy, statistical significance testing, and presentation quality. We are currently revising the manuscript and will address each comment in detail in a point-by-point response and revised manuscript.
  We particularly appreciate the referee's suggestions regarding clarification of the Relative Elevation Model (REM) and River Network Index (RNI), justification of flood-label thresholds, reporting of DeLong test statistics, and improvement of figures and language. These comments will substantially strengthen the manuscript.
  We thank the referee again for the constructive feedback and will carefully incorporate all recommendations in the revised version.
  
  Reply
  
  Citation: https://doi.org/10.5194/egusphere-2026-1275-AC2
RC2: 'Comment on egusphere-2026-1275', Anonymous Referee #2, 06 Jul 2026 reply

The manuscript proposes an ML framework for urban flood inundation mapping in the Mumbai Metropolitan Region, combining Sentinel-1, Sentinel-2, SRTM terrain attributes, and CHIRPS rainfall in Google Earth Engine. Two predictors presented as new - a Relative Elevation Model (REM) and a River Network Index (RNI) — are added to the feature stack, and flood labels are generated by consensus of SAR backscatter change, an NDWI threshold, and a REM threshold. A tuned XGBoost model is reported as the best performer, and an ablation experiment attributes a modest AUC gain to REM/RNI.
The topic is relevant, and the aim of a scalable, physically informed workflow for data-limited megacities is valuable. However, a structural circularity between label generation and model evaluation invalidates the reported skill, no independent validation is provided, the terrain predictors overlook the established literature on geomorphic flood descriptors, and the reference list contains duplicated and incorrect entries. These problems are foundational rather than presentational.
Major comments
Circularity between labels and predictors invalidates the evaluation. Flood labels are defined by the joint satisfaction of criteria based on SAR backscatter change, NDWI, and a REM threshold; the same variables are then used as predictors, and skill is computed against these self-derived labels. The model is thus rewarded for reproducing a function of its own inputs, not for detecting floods — the feature-importance figure confirms this, with REM by far the dominant feature and the SAR VV features ranking next. The ablation result — an AUC gain attributed to REM/RNI and presented as the paper's central finding — is a self-fulfilling artifact: removing REM necessarily degrades the prediction of labels partly defined by a REM threshold. Correcting this requires reference data independent of the predictor set and/or the exclusion of labeling variables from the feature stack.
No independent validation. The Mumbai monsoon floods of the study year are extensively documented (municipal flood-spot inventories, geolocated media reports, published flood maps, high-resolution imagery), yet none of these sources is used. The claimed "spatial congruency with known flood-prone regions" is purely qualitative. For a framework promoted as operational decision support, independent quantitative validation is indispensable, not future work.
Ground the terrain predictors in established geomorphic descriptors (HAND, GFI) and drop the novelty claim. The REM as defined is exactly the Height Above Nearest Drainage (HAND; Rennò et al.; Nobre et al.), used for well over a decade to delineate flood-prone areas. Geomorphic descriptors such as HAND and the Geomorphic Flood Index (GFI) have been specifically validated for flood-prone area delineation in data-scarce environments — precisely the manuscript's target setting — and would give the "hydrologic enhancement" a sound physical and bibliographic basis. I recommend that the authors reposition REM as an application of HAND and adopt, or benchmark against, the GFI or similar validated descriptors rather than ad hoc indices.
The RNI definition is internally contradictory. The text describes RNI as a Euclidean distance to the drainage network, but the formula given divides cumulative rainfall by the elevation above the regional minimum — it contains no distance term, is dimensionally arbitrary, and becomes singular near the coast. Please state what was actually computed, justify it physically, and rename it accordingly. Moreover, HydroRIVERS (coarse source hydrography) cannot represent Mumbai's urban drainage network; the adequacy of this river mask for REM/RNI computation must be demonstrated.
Note on the ongoing discussion
I share the concerns raised in the community comment (SAR double-bounce ambiguity in built-up areas; coarse working resolution for urban flood dynamics). The authors' reply does not resolve them: the argument that multi-source integration mitigates SAR artefacts fails here because the SAR features are themselves part of the label definition — errors propagate into the labels instead of being compensated; the reply lists "flow accumulation" among the predictors, but no such variable appears in the manuscript — this discrepancy should be reconciled; and the reply confirms a coarser working resolution than the Sentinel resolution emphasized in the manuscript — the actual common resolution must be stated explicitly.
Minor comments
The title is enclosed in quotation marks and mislabeled "Review article"; the byline uses "Research Scholar/Research Guide" instead of standard affiliations.
The study-area figure is cartographically inadequate, and the stated bounding box excludes part of the Mumbai Metropolitan Region. The flood-map figure uses pixel indices instead of coordinates, apparently classifies the sea, and the risk-zone panel is unexplained and nearly uniform. All maps need coordinates, scale, and a proper basemap.
The comparison table sets the proposed method against metrics from non-commensurable studies (different regions, events, validation data).
Duplicated in-text citations; bracket mismatches; the significance level should be denoted with the Greek alpha; the manuscript requires thorough professional English editing.

Reply

Citation: https://doi.org/10.5194/egusphere-2026-1275-RC2
RC3: 'Comment on egusphere-2026-1275', Anonymous Referee #3, 06 Jul 2026 reply

The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-1275/egusphere-2026-1275-RC3-supplement.pdf
Reply

Citation: https://doi.org/10.5194/egusphere-2026-1275-RC3
RC4: 'Comment on egusphere-2026-1275', Anonymous Referee #4, 08 Jul 2026 reply

While the authors are tackling an important issue, I have several concerns and questions about their methods, choice of datasets (e.g. ignored HAND, HAR, EAR) and the endless existing studies that use ML and DL for flood susceptibility mapping which use terrain descriptors and in many cases incorporate RS data which are not acknowledged in this work.
The Major issues I have with this manuscript are listed below:
1. literature review - there is a huge number of studies that largely do this (include hydrologic-topographic predictors) and test different ML models on a site. *e.g. search 'flood susceptibility'
2. REM - from a 30m DSM - which does not undergo any hydroconditioning. From what I understand you get some rivers from some place (undeclared) and compute vertical height. There are several models which do this, but also incorporate flow accumulation as well- making them much better substitutes, e.g. Height Above Nearest Drainage (HAND), height above river (HAR), elevation above river (EAR), The DEM is, I believe a DSM not a DTM.
3. RNI - this doesn't seem to even include river network or certain stream orders... the equation presented for RNI is cumulative precipitation divided by the different of the elevation at the pixel from some 'minimum regional' value. What is a region? How do the streams figure in this equation ? would a simple Euclidean distance raster from rivers greater or less than a stream order not suffice, if not, explain why.
these two 'new' datasets are not demonstrating they are better or more efficient than what is already available and how are they determined to be hydrologically significant predictors?
4. Classification of flood non-flood pixels. interesting approach to generate a consensus based label, but, are you confident this is working and the correct labels are being applied? Also, how did you decide that REM < 5 could/would demonstrate flooding? was there a sensitivity test or rationale behind this choice, please explain? same for NDWI. how did you come up with > 0.5 as the best threshold?
5. why is this described solely as an Urban model? could it not be used elsewhere?
6. table 1 misses the whole breadth of literature of ML models.
7. figure 1 - not impactful, what am I learning from this? maybe add terrain, land cover, something of greater interest to show the region
8. Modelling - many issues with this, starting with how labels are defined, but also you use REM to create labels and also in the model. ? this is a form of leakage and likely just means the model's performance does not represent its ability to predict real flood occurrence—only its ability to recover the rule used to create the label. How do you justify this decision?
9. results - auc, f1, etc. in some places you call it 'significantly superior and in the previous sentence/paragraph marginally better. Which is it?
10. it would be great to include a Known location with real flood extent to test your concept
11. figure 5 - I dont' find it convincing to support your confidence in this approach, also where did the low, med high thresholds for the right map come from? Suggest add points or areas of previous flooded monsoons could help strengthen narrative.
From the above list, I find the claims of this approach being an improved method and suitable method for flood hazard evaluation are not supported by the experiment and results.

Reply

Citation: https://doi.org/10.5194/egusphere-2026-1275-RC4
RC5: 'Comment on egusphere-2026-1275', Anonymous Referee #5, 10 Jul 2026 reply

The topic fits the scope of NHESS, but the study has a fatal methodological flaw, its results do not support its claims, and its presentation is below publishable standard. These problems cannot be fixed by revising the present analysis, so I recommend rejection. My reasons follow the three principal criteria.
The paper is submitted as a "Review article," but its content is a case-study research article, so the manuscript type is misassigned. Beyond this, the claimed novelty is overstated. The Relative Elevation Model is essentially the well-established HAND (Height Above Nearest Drainage) concept, yet it is presented as a new predictor without citing the relevant prior literature. The performance evidence also works against the authors' argument, since their own Table 5 shows the proposed method (71.7% accuracy, AUC 0.803) performing below the CNN/U-Net approaches they cite (75–85%, AUC 0.80–0.88), which contradicts the claim that the framework is competitive.
The most serious problem is a target-leakage issue that undermines the paper's main result. The flood labels are defined in part by REM < 5 m (Sect. 4.3), but REM is then fed back in as a predictor, and the ablation in Table 4 attributes a 5–6% AUC gain to REM/RNI and treats this as the key evidence of novelty. Since the model is essentially predicting labels built from one of its own inputs, that gain reflects circularity rather than genuine predictive skill. The RNI predictor is also internally inconsistent, as Sect. 4.2 describes it as a Euclidean distance to the drainage network while the accompanying formula, RNI = P(x,y) / (DEM − DEM_min), is precipitation divided by relative elevation, which has nothing to do with distance and is dimensionally incoherent. Compounding these issues, there is no independent validation. Both training and testing rely on the same consensus pseudo-labels, so the reported accuracy really measures agreement with synthetic labels rather than observed flooding. The statistical claims are similarly thin, with the AUC difference between models (0.796 vs 0.803) repeatedly described as significant even though no p-values or confidence intervals are reported. Finally, reproducibility is only partial. The data are commendably shared on Zenodo, but the Google Earth Engine scripts, model code, and hyperparameter ranges are not.
The English requires thorough editing throughout, as there are numerous awkward phrasings. The text is also repetitive. the Introduction and Literature Review restate the same point — that ML methods ignore hydrologic–topographic controls — four or more times, and these two sections should be merged and condensed. The references are inconsistently formatted, and the equations are poorly rendered, with broken subscript alignment.

Reply

Citation: https://doi.org/10.5194/egusphere-2026-1275-RC5

Ankush S. Pawar and Gayatri M. Phade

Data sets

Data for Hydrologic–Topographic Enhanced Machine Learning for Urban Flood Inundation Mapping Ankush S. Pawar and Gayatri M. Phade https://doi.org/10.5281/zenodo.18486214

Ankush S. Pawar and Gayatri M. Phade

Viewed

Total article views: 417 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
304	84	29	417	23	22

HTML: 304
PDF: 84
XML: 29
Total: 417
BibTeX: 23
EndNote: 22

Views and downloads (calculated since 22 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	155	45	18	218
May 2026	111	18	1	130
Jun 2026	26	6	6	38
Jul 2026	12	15	4	31

Cumulative views and downloads (calculated since 22 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	155	45	18	218
May 2026	111	18	1	130
Jun 2026	26	6	6	38
Jul 2026	12	15	4	31

Viewed (geographical distribution)

Total article views: 387 (including HTML, PDF, and XML) Thereof 387 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 16 Jul 2026

Short summary

Urban floods frequently affect coastal cities like Mumbai, causing severe damage. This study develops a satellite-based method that combines rainfall, terrain, and radar data with machine learning to map flood areas more accurately. By including terrain features that influence how water spreads, the model improves flood detection reliability. The approach provides a scalable and practical tool for supporting flood monitoring and decision-making in rapidly growing cities.


Total:	0
HTML:	0
PDF:	0
XML:	0