Pluvial flooding damage scenarios using remote sensing and deep learning models for urban object detection

Cauteruccio, Arianna; Maragliano, Matteo; Rajabi Toostani, Roozbeh; Boni, Giorgio; Moser, Gabriele

doi:10.5194/egusphere-2026-1625

Preprints

https://doi.org/10.5194/egusphere-2026-1625

Preprints

19 May 2026

| 19 May 2026

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

Pluvial flooding damage scenarios using remote sensing and deep learning models for urban object detection

Arianna Cauteruccio, Matteo Maragliano, Roozbeh Rajabi Toostani, Giorgio Boni, and Gabriele Moser

Abstract. A quantitative assessment of the risk of pluvial flooding in an urban area is subject to various uncertainties. These include evaluating the intensity and patterns of rainfall across the area, and the exposure and specific vulnerability of different targets. In this study, the impact of rainfall pattern and intensity description is evaluated using rainfall information obtained from innovative opportunistic sensors based on satellite links and traditional rain gauges, located both within and outside the flood-prone area. Urban Flood Drifters (UFDs) are considered the primary targets, with a focus on motor vehicles (cars, vans, and motorbikes). Aerial images and deep learning–based object detection are used to evaluate the exposure of potential UFDs in terms of their numerosity and location within flood-prone urban areas. The specific vulnerability of these targets is considered using stability and damage curves from the literature, based on water depth and flow velocity. The magnitude and spatial distribution of these flood characteristics are obtained by running hydrodynamic simulations of two-dimensional pluvial flooding propagation in the urban area, based on rainfall patterns obtained from both opportunistic and traditional sensors. The impact is demonstrated in terms of vehicle safety, monetary damage experienced by flooded vehicles, and traffic disruption using an observed event. The results obtained from different rainfall data sources and aerial images with different spatial resolutions are compared and discussed in terms of their suitability for object detection and risk assessment. Flood hazard maps demonstrated that rain gauges positioned outside of the study area do not capture the actual magnitude of the event measured at the reference station. On the contrary, opportunistic sensors, despite significant underestimation of the peak rainfall, provide a good representation of the temporal evolution and total volume of the reference rainfall event. Damage scenarios demonstrated that the role of false negatives, in automated object detection, can be relevant to the safety analysis and the assessment of monetary damage, especially when spatial resolution is coarser, while it is quite negligible for the assessment of traffic disruption.

Received: 23 Mar 2026 – Discussion started: 19 May 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

Arianna Cauteruccio, Matteo Maragliano, Roozbeh Rajabi Toostani, Giorgio Boni, and Gabriele Moser

Status: open (until 03 Dec 2026)

Post a comment Subscribe to comment alert

CC1:
'Comment on egusphere-2026-1625', xing wang, 24 Jul 2026 reply
The idea of comparing rainfall inputs from satellitelink opportunistic sensors (SRS) with conventional rain gauges, and of coupling these with a YOLObased object detector to map vehicle exposure, is certainly appealing. The threetier damage assessment—specific energy for safety, depthdamage curves for monetary loss, and depthvelocity reduction for traffic disruption—provides a coherent framework that could be transferable to other cities. The observation that false negatives affect safety and economic estimates more than traffic disruption is also a useful nuance. However, these positive aspects are overshadowed by a number of fundamental issues that need to be addressed.
Major concerns
The most critical problem is the deliberate neglect of the stormwater drainage network, including inlets (Section 2.2). Pluvial flooding is by definition a consequence of rainfall exceeding the system’s conveyance capacity. By removing this capacity entirely, the model essentially turns the study area into a closed basin where all rainfall accumulates on the surface. This systematically overestimates water depths and flooded extents, and more importantly, it may artificially favour the SRSbased results. Under nodrainage conditions, flood volume is nearly a linear function of total rainfall—which the SRS happens to capture reasonably well. If drainage were included, the peak rainfall intensity (which the SRS underestimates significantly) would become much more influential in determining peak water levels. The conclusion that SRS outperforms distant rain gauges could therefore be a direct artefact of the missing drainage. At a minimum, I would expect a sensitivity test with a simplified drainage representation to quantify this effect.

The flood simulation is tied to the 24 September 2022 event, yet the aerial images used for vehicle detection are not shown to have been acquired on that day, let alone during the peak rainfall period (around 09:15–09:45). The authors admit this as an assumption (line 496), but without any validation for this particular event, the exercise amounts to overlaying a static vehicle inventory from an ordinary day onto a dynamic flood scenario. This renders the absolute monetary loss estimate—roughly 4 million euros—essentially unverifiable. If eventday imagery is unavailable, the study should be reframed explicitly as a hypothetical scenario demonstration rather than a postevent impact assessment.

The authors argue that the SRS captures total rainfall well but underestimates the peak because the satellite link crosses the rain core while the sensor itself is not directly under the core. This reasoning is not physically convincing: if the link indeed traverses the rain core, the pathintegrated attenuation should reflect the high intensities along that segment. That the system consistently smooths the peak instead suggests a fundamental limitation of the retrieval algorithm—the uniformrainfall assumption over the link length, which is known to fail under convective, shortduration storms. The paper would benefit from a more candid discussion of this algorithmic uncertainty and its implications for flood modelling.

A key takehome message (lines 484–489) is that spatial variability of rainfall matters more than objectdetection classification errors. To substantiate this claim, one would need controlled comparisons where either rainfall or detection quality is varied while the other is held fixed. Here, however, the comparison of rainfall inputs uses the highresolution (5 cm) detection, while the comparison of detection performance confounds resolution changes (5 cm vs. 20 cm) with the rainfall source. This means two variables are altered simultaneously, making it impossible to attribute the observed differences to one factor alone. The authors should either redesign the analysis to isolate these effects or substantially tone down their conclusion.

Safety is assessed via the Kramer specificenergy approach, which couples depth and velocity, whereas monetary loss and traffic disruption are based on depthonly curves. For pluvial floods, flow velocities are generally low, so the velocity term in specific energy contributes little. This creates a situation where a site with, say, 0.35 m depth and very low velocity would fall into a “medium” specificenergy class but incur near100% monetary loss—a clear inconsistency. The authors need to either justify why this discrepancy is acceptable or adopt a unified metric across all impact categories.

Table 4 shows an F1 score of only 0.68 for motorcycles, with false negatives of 35% and false positives of nearly 30%. At this level, the model is effectively confusing motorcycles with shadows, manhole covers, and vegetation. Including such noisy data in the monetary loss calculations undermines confidence in the corresponding figures. I would recommend treating twowheeler results as a proofofconcept only, or improving the detection model specifically for small objects before integrating them into the risk assessment.

The analysis involves a chain of uncertainties: SRS rainfall retrieval (peak errors >50%), hydrodynamic discretisation (5 m grid, 0.5 s timestep), object detection (FN 5–35%), and damage curve fitting. Yet the final monetary loss is reported as a single deterministic number without any confidence interval or sensitivity range. For an engineering risk assessment, this is insufficient. Even a simple firstorder error propagation or a scenariobased interval would greatly improve the credibility of the results.

Minor issues
The abstract currently fails to mention the paper’s most striking conclusion—that rainfall variability is more influential than detection errors. This should be added to help readers grasp the core message from the outset.
False positives (FP) receive too little attention. At 20 cm resolution, FP for fourwheelers reaches 10%, meaning about 350 detected “vehicles” are actually on rooftops or trees. While these may not affect flooddamage calculations directly, they indicate a poor semantic understanding of urban scenes, and their spatial distribution could mislead exposure mapping.
The Manning coefficient for impervious surfaces (0.018) is adopted without local justification. Given that urban surfaces vary widely (asphalt, concrete, paving stones), a sensitivity analysis would be prudent.
Radar data are presented in Figure 4 but never used as an input in the modelling. Since the paper focuses on comparing rainfall sources, the exclusion of radar—a common urban monitoring tool—seems arbitrary and should be explained.
The assumed 30% share of minivans, uniformly distributed, is not backed by local vehicle registration data. This proportion directly affects the monetary loss estimates, and a sensitivity test or actual traffic statistics would strengthen the analysis.
The filtering rule for moving vehicles—using a 1–2 m buffer along main streets—is crude and unvalidated. Some roadside parked vehicles also affect traffic, and some moving vehicles may have been evacuated before the flood. The authors should either validate this rule or acknowledge its limitations more explicitly.
Finally, the data availability statement indicates that key georeferenced layers are available “upon request,” rather than deposited in a public repository. To support reproducibility—especially for a methodoriented paper—I strongly encourage the authors to upload their intermediate datasets (annotations, flood maps) to an open archive with a DOI.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-1625-CC1
RC1:
'Comment on egusphere-2026-1625', Anonymous Referee #1, 29 Oct 2026 reply
This manuscript presents an integrated framework for assessing the urban pluvial flooding impacts in the Metropolitan area of Genoa, Italy, using a rainfall event that occurred on 24 September 2022 as a case study. The framework combines rainfall measurements from conventional rain gauges and innovative microwave sensors, two-dimensional hydrodynamic modelling with HEC-RAS, vehicle detection from aerial imagery using YOLOv8n and YOLOv11n, and vehicle-specific damage functions. The simulated water-depth and velocity fields are combined with detected vehicle locations to estimate vehicle safety, monetary damage, and traffic disruption. The manuscript also examines how the choice of rainfall data source, image spatial resolution, and object-detection errors affect the resulting flood-damage scenarios.
General Comments:
This study addresses a relevant issue in urban flood risk assessment by demonstrating a potentially useful integration framework, and the proposed workflow has practical value. However, the novelty and significance of the study is not particularly strong because several methodological assumptions require further justification before the quantitative findings and broader conclusions can be considered sufficiently robust.
The comparison between rain gauge and the Smart Rainfall System (SRS) is potentially insightful, but the current experimental design does not isolate the effect of sensor type. A rain gauge represents a point measurement, whereas the SRS provides rainfall attenuation along the extended satellite link. Although Arpal-FI is treated as the reference, a point measurement near the maximum of a localized convective event may not necessarily represent catchment-wide rainfall. Meanwhile, the SRS differ in location and in their intersection with the rainfall field. Therefore, applying each time series uniformly across the model domain combines the effects of senor technology, sensor location, and storm geometry. It would be appreciated if the authors could clarify how the rainfall time series were spatially applied and distinguish measurement accuracy from event-specific spatial representativeness.

The hydrodynamic model is not validated against observations of flood extent, flood depth, road closure, or other evidence from the September 2022 event. This makes it difficult to determine whether the simulated flood characteristics represent the actual event. Also, roof runoff is assumed to be completely discharged into the underground drainage network, whereas street inlets and the remaining drainage system are neglected. This assumption requires further justification because it may substantially affect flood depth, duration, and the estimated monetary damage. It would be appreciated if the authors could validate the hydraulic model and evaluate its sensitivity to drainage capacity, surface roughness, mesh resolution, boundary conditions, and terrain processing.

The deep-learning analysis requires more information about training, validation, and independent testing. It is unclear whether the training, validation, and whole-area evaluation samples are spatially independent. The comparison between the 5cm and 20cm imagery is also confounded by differences in acquisition time, traffic conditions, lighting, and possibly sensor characteristics. As a result, the reduction in detection performance cannot be attributed to spatial resolution alone.

The exposure and damage scenarios involve several assumptions that are not adequately documented. The aerial images do not appear to have been acquired during the September 2022 flood event and should not be regarded as the actual vehicle distribution during the event. The classification of the two datasets as rush-hour and off-peak conditions also requires exact acquisition dates and times. Similarly, a static aerial image cannot directly determine whether a vehicle is moving. The method used to classify vehicles as moving or parked based on their position within a road strip should be explained and validated.

The monetary-damage assessment requires greater transparency and uncertainty analysis. The manuscript does not provide the assumed vehicle values, valuation year, commercial sources, or separate damage totals for cars, minivans, and motorcycles. The assumption that minivans represent 30% of four-wheeled vehicles and are uniformly distributed across the study area is also not justified. The motorcycle damage curve is newly constructed from assumed component thresholds and commercial costs, but the underlying component information and cost data are not presented. It would be appreciated if the authors could provide these data, define the valid ranges of all damage functions, and quantify how uncertainty in the damage curves, vehicle values, vehicle composition, and hydraulic results affects the reported damage.

Specific Comments:
L47-51: It would be useful to explain more explicitly how the present study advances beyond the direct spatial combination of existing stability criteria and flood maps.
L77-103: The descriptions of the general YOLO model could be shortened, while more information could be provided about the actual model configuration, pretrained weights, optimization settings, augmentation, and IoU threshold used in this study in section 2.3.
L145-153: the objectives are described mainly as a sequence of methodological steps. It would be appreciated if the authors could re-write and reformulate them as explicit research questions.
L165-195: Please provide the exact locations or coordinates of the rainfall instruments and explain how measurements with different spatial supports were converted into rain-on-grid inputs. Were the measurements applied uniformly to every computational cell?
L190-195 and L319-328: Meteorological radar observations are presented to illustrate the event evolution, but they are not included as a rainfall-forcing scenario. Could the authors explain why radar rainfall was excluded from the hydrodynamic comparison?
L209-220: The omission of the drainage network is justified by the insufficient number and nonoptimal distribution of inlets reported in a previous study. However, an insufficient drainage system can still remove a substantial volume of runoff. Please quantify the implications of assuming zero street-drainage capacity.
L222-237: Please provide information about hydraulic model calibration and validation. Were any observations of inundation extent, water depth, flood duration, road closure, or reported damage available for the September 2022 event?
L227-233: Upsampling the 1m DTM to 0.25m does not add new terrain information, while Gaussian smoothing may alter small-scale urban flow controls such as curbs and local depressions. It would be appreciated if the authors could justify this processing and evaluate its influence on the results.
L239-253: Please provide the acquisition dates, approximate acquisition times, platforms, sensors, spectral information, and processing methods for the 5cm and 20cm aerial images. Please also explain whether each dataset represents a single acquisition time or an orthomosaic assembled over an extended period.
L247-253: It is unclear what distinguish the “traditional annotation” from the “detailed annotation” approach. Table 3 reports the same numbers of training and validation instances for both approaches. Were the same vehicles reannotated, or were additional visibility conditions introduced without changing the number of objects? Please clarify.
L360-368: The comparison between YOLOv8n and YOLOv11n is based on only five validation patches containing 179 instances. This appears insufficient to conclude that YOLOv11n offers no improvement. Please report variability across repeated training runs or qualify this comparison as preliminary.
L379-390: Please clarify whether the evaluation across the entire study area includes patches or locations used during training and validation. If so, it should not be described as a test-stage evaluation.
L455-473: The method used to distinguish moving from parked vehicles is unclear. Please define the road strips precisely and explain how the classification was validated. The resulting output represents local depth-constrained vehicle speed rather than network-level traffic disruption, which would require consideration of connectivity, rerouting, and travel-time effects.
Remarks:
L47: Please remove the duplicated full stop after “flooding.”
L64-69: Redmon et al. (2016) introduced YOLO and does not appear to be the appropriate reference for Faster R-CNN. Please check.
L270-275: The low specific-energy class is defined as E ∈ [0.1m, 0.3m]. How are flooded locations with E<0.1 m classified?
L395: Please remove the additional “o” in “5 cm o is provided.”
Figure 11: The caption refers only to the 5 cm resolution dataset, whereas the figure and surrounding text appear to discuss both 5 and 20 cm datasets. Please revise the caption.
Table 4: Precision and recall appear to be reversed relative to Equations (1) and (2). Please verify the calculations, denominators used for TP, FN, and FP percentages, and the resulting F1 scores.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-1625-RC1

Arianna Cauteruccio, Matteo Maragliano, Roozbeh Rajabi Toostani, Giorgio Boni, and Gabriele Moser

Viewed

Total article views: 506 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
302	154	50	506	40	33

HTML: 302
PDF: 154
XML: 50
Total: 506
BibTeX: 40
EndNote: 33

Views and downloads (calculated since 19 May 2026)

Month	HTML	PDF	XML	Total
May 2026	117	34	7	158
Jun 2026	16	6	3	25
Jul 2026	38	16	4	58
Aug 2026	6	12	4	22
Sep 2026	39	23	11	73
Oct 2026	72	45	16	133
Nov 2026	14	18	5	37

Cumulative views and downloads (calculated since 19 May 2026)

Month	HTML	PDF	XML	Total
May 2026	117	34	7	158
Jun 2026	16	6	3	25
Jul 2026	38	16	4	58
Aug 2026	6	12	4	22
Sep 2026	39	23	11	73
Oct 2026	72	45	16	133
Nov 2026	14	18	5	37

Viewed (geographical distribution)

Total article views: 492 (including HTML, PDF, and XML) Thereof 492 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 11 Nov 2026

Short summary

The risk of pluvial flooding in an urban area is addressed using a multidisciplinary approach involving hydrodynamic simulation and deep learning–based object detection to evaluate the exposed motor vehicles in the flooded areas. The objective is to quantify the potential damage based on rainfall patterns obtained from both opportunistic and traditional sensors. The impact is demonstrated in terms of vehicle safety, monetary damage experienced by flooded vehicles, and traffic disruption.


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