https://orcid.org/0000-0003-4390-2664
Figures
Abstract
Flooding in urban areas causes significant economic and social impacts on populations across the globe. Flood detection plays a pivotal role in disaster management, necessitating advanced methodologies to enhance accuracy and efficiency. Addressing this challenge requires delineating flood extent at a high spatial and temporal resolution. Efforts to fully quantify urban flood distribution utilizing the potential of Synthetic Aperture Radar (SAR) imageries in a cloud-based platform have ample potential but have yet to produce viable results in the urban landscape. Flood detection has been a challenging task in urban areas due to limitations of spatial-temporal resolution and complex back scatter mechanisms in urban settings. However, advancement in big-data and cloud-computing, data acquisition, satellite image processing and predictive analysis are rapidly becoming more accessible. Building on recent advancements, this study presents an analysis of methods exploring and comparing identification of flooded areas in urban and peri-urban locations, which has not been fully described. Using Houston, TX to test these methods, we compare flood maps generated from multiple classification method including constant threshold Change Detection Approach (CDA), Otsu method, and Machine Learning (ML) classification with Random Forest (RF) model using Sentinel-1 SAR images in Google Earth Engine (GEE). An extensive performance evaluation is conducted, including accuracy assessments, precision, recall, F1-score, and confusion matrices. The CDA approach shows the highest accuracy in peri-urban areas, while ML classifier outperforms both CDA and Otsu in urban settings. The analysis in this paper contributes to the development of flood detection methodologies in support of urban flood management.
Citation: Al Mehedi MA, Smith V, Kremer P (2025) A comparative analysis of urban and peri-urban flood identification using SAR imagery. PLOS Water 4(9): e0000269. https://doi.org/10.1371/journal.pwat.0000269
Editor: João Miguel Dias, Universidade de Aveiro, PORTUGAL
Received: May 15, 2024; Accepted: August 7, 2025; Published: September 11, 2025
Copyright: © 2025 Al Mehedi et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All data used in this study are publicly available and freely accessible. The primary datasets utilized include: 1. COPERNICUS Sentinel-1 SAR GRD: This dataset is publicly available through Google Earth Engine and can be accessed at https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S1_GRD. 2. National Land Cover Database (NLCD) 2019: This dataset is also publicly available through Google Earth Engine and can be accessed at https://developers.google.com/earth-engine/datasets/catalog/USGS_NLCD_RELEASES_2019_REL_NLCD. No legal or ethical restrictions are imposed on the availability of this data. Fellow researchers can access these datasets directly through the provided links.
Funding: This work was supported by the Department of Civil and Environmental Engineering, Villanova College of Engineering, and Villanova University (support to MAAM, VS, and PK). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
1 Introduction
Flooding is the most expensive natural catastrophe, costing billions worldwide in property as well as loss of life [1–5]. Due to climate change and urban growth, the frequency and unpredictability of flood events, and the impacts of floods have increased in recent times. At the same time, urban growth results in an increase in impermeable surfaces in urban areas [6,7]. Climate change projections suggest that the global population affected by floods will double by 2030, rising from the current 72 million people per year to an estimated 147 million [8–10]. Watershed development associated with urbanization causes greater urban runoff and a higher rise of flash floods. These challenges have compounded, making urban flooding a major challenge for cities around the world.
Despite its profound impact, urban flooding is often difficult to define. A lack of flood extent data makes it challenging to identify areas of risk and vulnerability. Accurate flood extent mapping is critical for flood management, minimizing losses, supporting emergency response and recovery, and improving flood prediction models for informed decision-making. Accurately mapping the extent of floods is crucial for effective flood management, enhancing the efficiency of emergency response and recovery efforts and providing essential data to improve flood prediction models for better-informed decision-making, ultimately minimizing economic and human losses. Observation-driven flood monitoring techniques, such as satellite imagery analysis, ground-based sensors, and drone surveillance, have become essential due to the severe uncertainty associated with commonly used flood risk models, particularly in regions vulnerable to climate change-induced flooding [11,12]. Networks of observational sensors can be limited [13]. To fill this need, monitoring systems are increasingly incorporating satellite-based Earth Observation (EO) data, providing synoptic and repeated views of potentially inundated areas [14–16]. Researchers have been exploring flood mapping and monitoring methods to provide real-time, near-real-time, and accurate flood information [17–21]. However, conventional flood mapping using optical imagery, such as Landsat and MODIS, is limited by its inability to penetrate through cloud cover [22,23]. Therefore, during natural disasters, such as major storms, it is often difficult to collect data. This is particularly true for mapping flooded areas under adverse weather conditions [24,25]. For instance, spectral indices like Modified Normalized Difference Water Index (MNDWI) and Normalized Difference Water Index (NDWI) derived from optical images can show flood extent, but often have significant limitations due to cloud coverage [26–30].
EO data from satellites, such as Sentinel-1 Synthetic Aperture Radar (SAR), has enabled comprehensive and repetitive views of flooded regions and is increasingly being utilized in disaster monitoring systems. Radar sensors, like Sentinel-1 SAR, overcome limitations of optical data by providing high quality images regardless of weather conditions or time of the day [31]. SAR can also penetrate partially or fully through vegetation canopy, enabling flood detection in forested regions, where optical images may be insufficient [32–34]. While optical data is highly effective for mapping water areas in cloud-free conditions, SAR imagery proves valuable in mapping flood extent regardless of weather conditions common to major storms [35–37].
SAR data has been proven to be an effective tool for mapping floodwaters in vegetation-free areas, where distinct backscattering signatures of bodies of water are observed [38–41]. With the advancement of automatic algorithms, SAR data is now well-established in practical flood control, enabling near-real-time floodwater mapping [42–45]. However, the detection of flooding in complex environments such as urban regions remain challenging, as radar signatures in such areas can be ambiguous and hard to predict [46–48]. It is difficult to use SAR data to distinguish between water and water-like surfaces, especially in urban areas with different types of reflective surfaces and impervious smooth surfaces [49,50]. In urban areas, SAR-based flood detection also faces difficulties due to the presence of complex infrastructures and heterogeneous surfaces, leading to challenges in accurately delineating flooded areas from other features [51–53]. Nonetheless, it is important to overcome these barriers to enable wide use of SAR imagery like Sentinel-1 for mapping flood extent for in large urban areas due to its free availability, medium resolution, and availability during any weather conditions [54,55].
SAR based operational flood mapping services primarily focus on rural areas because of the challenges of using SAR in urban landscapes. A few studies showed that flooding can be mapped in urban environments using bi-temporal change detection and advanced methods like ML models using SAR. It has been demonstrated that utilizing a 10m-resolution, bi-temporal SAR can be efficient for detecting floods in urban environment [32,52,56,57]. Unlike rural areas where SAR signals typically interact with relatively uniform surfaces such as open water or vegetation, urban landscapes are composed of highly heterogeneous materials and structures. The presence of vertical buildings, narrow streets, and complex infrastructure introduces intricate scattering effects—such as double-bounce reflections, corner scattering, and shadowing—which produce mixed or ambiguous backscatter signatures [58]. These effects complicate the accurate identification of inundated surfaces, particularly when smooth impervious materials (e.g., roads or pavements) mimic the low backscatter of open water, or when vertical features distort signal return [59]. These complexities highlight the necessity for developing and applying specialized techniques that can handle the unique SAR signal behavior in urban flood mapping. The backscatter characteristics of ground objects in SAR images play a crucial role in distinguishing between different objects and detecting changes in backscatter properties [60–62]. Various techniques utilized with SAR images for flood mapping include histogram threshold methods, RGB composition, change detection methods, data fusion with optical images, and classification techniques, including unsupervised and supervised approaches [63–66]. In tandem with these established techniques, constant threshold Change Detection Approach (CDA) with the ratio image technique emerges as a valuable method in urban flood mapping using SAR data. This approach involves establishing a consistent threshold for backscatter changes and employing a ratio image technique to enhance the discrimination of flood-affected areas [33,67]. From previous research, it is evident that ML classification methods show promise in urban flood mapping by autonomously learning complex patterns from SAR imagery, contributing to improved accuracy in flood extent delineation [68,69].
The complexity of the backscatter mechanism in urban and peri-urban areas poses unique challenges in accurately mapping floods compared to regions dominated by simpler land cover types [70–72]. The intricate mix of surfaces, structures, and materials in urban and peri-urban environments can result in increased uncertainties and errors in flood mapping. This underscores the necessity of dedicated research and tailored methodologies to overcome the inherent complexities and accurately quantify flood extents in these specific land cover types [70,73–75].
The applicability of Google Earth Engine (GEE) in flood mapping for urban areas is highly promising, leveraging satellite image time series such as Landsat to create global-scale data products [44,76–78]. The GEE computing platform enables computationally efficient investigations of land surface dynamics. Notably, GEE’s repository includes regularly updated Sentinel-1 Ground Range Detected (GRD) data, overcoming the challenges of SAR preprocessing and facilitating broader adoption of SAR remote sensing for flood mapping purposes [79–81].
This study addresses three key challenges: comparing SAR-based flood detection methods across urban and peri-urban areas during a real-world event; assessing the impact of urban surface complexity on detection performance; and enabling rapid flood mapping using publicly available Sentinel-1 data GEE for quick decision-making. This study analyzes the performance of flood detection methodologies customized for urban and peri-urban land cover, by comparing and leveraging using publicly available SAR data and GEE computational capabilities for flood mapping. Additionally, we assess and compare two flood detection methodologies, CDA and ML classification. Key performance metrics, including the components of confusion matrix, precision, recall, and F1-score, are analyzed to offer a comprehensive understanding of the flood detection capabilities of the CDA and ML classifier. This study provides quantitative evidence of the accuracy and reliability of each approach in identifying flooded locations. While more advanced methods exist, our goal is to evaluate practical, rapid solutions for flood detection. This analysis illuminates and enhances the capability for interpreting the high range of noise and heterogeneous urban SAR backscatter signatures for flood delineation. This comprehensive assessment provides valuable insights into urban flood mapping, emphasizing the significance of tailored methodologies for urban and peri-urban landscapes, which ultimately enhances flood risk assessments and disaster management in urbanized areas.
2 Data and methods
2.1 Study area
Houston is one of the largest cities in the US, with a population of 2.29 million and a land area of roughly 1740 km2 [82]. Due to its location on the gulf coast, it is frequently subjected to damaging hurricanes. Houston experienced considerable destruction when Hurricane Harvey hit Texas between August and September 2017 [83,84]. Hurricane Harvey, a Category 4 storm, made its initial impact on San Jose Island, Texas, on August 25, 2017. During this event an unprecedented volume of rainfall accumulated across Eastern Texas and Louisiana over a span of four days. Following its landfall on the Texas mainland on the same day, Harvey weakened into a tropical storm [85–87]. The slow movement of the storm resulted in the region receiving exceptional volumes of rainfall. The result was flooding that took days to drain via rivers, leading to widespread catastrophic flooding, particularly in the Houston metropolitan area. The local National Weather Service office in Houston recorded daily rainfall accumulations of approximately 370 mm (14.57 inches) on August 26 and 408 mm (16.08 inches) on August 27. Hurricane Harvey has been called the wettest However, given the detailed on record in the United States [88–90]. The catastrophic storm caused a total economic loss of nearly US$125 billion, forced more than 30,000 people to leave their homes, triggered more than 17,000 rescue operations, and resulted in 106 fatalities in the United States [91].
2.2 Data and image
In this study, we analyzed Sentinel-1 SAR images through the GEE platform to compare the quality of flood extent delineation using methods that are near real time and methods that require more extensive processing. Sentinel-1 SAR data, is a publicly accessible radar dataset, with a resolution of 10 meters, that enables comprehensive spatial analysis, particularly in the context of large-scale flood mapping [36,92]. The Sentinel 1 GRD (Ground Range Detected) products available on GEE platform (https://developers.google.com/earth-engine/datasets/catalog/sentinel) are fully preprocessed and ready for analysis. This dataset encompasses data from both cross-polarized (VH) and like-polarized (VV) channels in the form of Level-1 GRD products. Our research has harnessed VV polarizations from this GRD product for effective flood mapping. VV polarization was used due to its consistent availability in GEE and reliable sensitivity to urban flood signals, particularly from double-bounce effects [12]. This also ensured computational efficiency for near-real-time mapping [21]. While VH and coherence data can improve detection in complex environments, they were excluded here for simplicity. Future work will explore their integration. This study focuses on detecting floods in open urban and peri-urban areas, where SAR backscatter is more reliable. Densely built-up zones were not emphasized due to complex scattering effects and potential signal ambiguity. Future studies should address these challenges for comprehensive urban flood mapping. Ascending orbit was selected for this study.
Sentinel-1 imagery along with satellite imagery, and hydrographic data (surface water bodies) from August 29th, 2017, were retrieved through GEE. We used a Digital Elevation Model (DEM) from WWF HydroSHEDS with 30-m spatial resolution, which was compatible with the Sentinel-1 image and obtained from the USGS within the GEE data catalog [93]. Additionally, we utilized the Joint Research Centre (JRC) Global Surface Water (GSW) layer, available within the GEE platform, to identify permanent and semi-permanent water bodies within the study area [94]. The Dartmouth Flood Observatory (DFO) map is used with the assistance of High-Water Marks (HWM) dataset to identify flooded locations [79], as a benchmark for assessing the performance of the models under investigation [95,96].
The SAR reflectance characteristics of land cover types vary significantly, with features like forests and water bodies displaying lower reflectance compared to urban areas [97,98]. To evaluate the magnitude of flooding in both densely populated urban areas and peri-urban areas with more open spaces and less impervious areas, the National Land Cover Database (NLCD) 2021 is employed. The NLCD is a 30-m spatial resolution Landsat-based land cover database. The classification relies on the imperviousness data layer for urban classes and a decision-tree classification based on LANDAST imagery for the remaining classes. NLCD products are generated by the Multi-Resolution Land Characteristics (MRLC) Consortium, a collaboration among federal agencies led by the U.S. Geological Survey [99]. Land cover class 24, which represents highly developed areas with significant residential or occupational occupancy, such as apartment complexes, row houses, and commercial/industrial sites, was used to identify dense urban areas (Fig 1). Impervious surfaces account for 80% to 100% of the total land cover in this class. To identify peri-urban areas, land cover class 21 was selected, which encompasses areas characterized by a mixture of constructed materials and predominantly vegetated areas with lawn grasses. Impervious surfaces make up a minority of the overall land cover, accounting for under 20% in these areas. Peri-urban areas often encompass various features such as large-lot individual homes, parkland, golf courses, and cultivated greenery found within urbanized areas. These areas serve various functions, including leisure activities, erosion prevention, and enhancing visual appeal.
The base map layer was obtained from ArcGIS Basemap, and the administrative boundary shapefile was sourced from the City of Houston GIS (COHGIS) Open Data Portal (COH Administrative Boundaries dataset, https://cohgismycity.opendata.arcgis.com/).
2.3 Preprocessing and refinement
Many of the preprocessing tasks needed for analyzing SAR images have already been completed for the images available in the GEE data catalog [100]. First, we determined the study area, the city of Houston. A shape file of the Houston administrative areas was used in the GEE platform to locate Sentinel-1 images within the designated study area. Second, the images were filtered from the Sentinel-1 collection based on dates during flood and dry periods. Third, the collection was filtered based on additional criteria. The instrument mode was filtered, to select the Interferometric Wide (IW) swath mode. Following the orbit direction is filtered, VV polarization option is selected. Next, a speckle filter was applied to the Sentinel-1 images to enhance flood mapping accuracy. Speckle filters and been shown to effectively reduce noise, and improve the clarity of flooded area detection, contributing to more reliable flood assessment during disaster events [80,101]. A Refined Lee speckle filter [102], adapted for Sentinel-1 SAR images, was applied to reduce noise and improve flood mapping accuracy. This filter preserves edges while smoothing homogeneous areas, making it suitable for urban flood detection.
Consequently, two filtered image collections were created for the study area: one comprising image captured in a dry period prior to the flood event and another consisting of images captured during the flooding period. For the dry period, the date range was set from August 1, 2017, to August 15, 2017, resulting in the identification of two images that meet all the specified criteria. For the flood event images, we applied filtering between August 25, 2017, and August 31, 2017, resulting in the identification of one image that met the specified criteria. The image was taken on August 29, closely corresponding to the primary impact date of August 25. Three refinement methods were applied to avoid misclassification, as recommended by UN-SPIDER [103]. The first refinement involved the exclusion of permanent and semi-permanent water bodies, which were identified as areas containing water for more than 10 months using the JRC GSW [104]. These areas were removed to eliminate any false-positive flooded regions. Subsequently, areas with a slope greater than 5%, were eliminated. This criterion was based on the understanding that pixels with slopes higher than this threshold would not retain water and would drain towards lower elevation areas. To perform this refinement, the WWF Hydrosheds, an elevation model available within the GEE catalog, was utilized [93]. After generating the initial flood map, we removed regions labeled as flooded if they comprised fewer than eight connected pixels for further refinement. These small areas, with an area of less than 6400 square meters, were excluded from the definition of flooded regions for this study. After applying these three refinements, we produced the ultimate flood map.
2.4 Change detection approach with constant threshold
The change detection approach relies on a pixel-based comparison of the difference in reflection between the dry and flooded images [105,106]. Fig 2 presents a comprehensive workflow detailing the Change Detection Approach (CDA) with data preprocessing. Initially, a division image is computed from Sentinel-1 imagery of dry and flood period. The distribution of the ratio values is presented in S2 Fig. Subsequently, a range of constant threshold values (ranging from 1.01 to 1.05) are applied to the division image to identify areas affected by flooding for the CDA [33]. Increasing the constant threshold in flood mapping may lead to more false negative flooded locations due to its impact on the detection method’s sensitivity. When the threshold is raised, the algorithm becomes less permissive in identifying changes or fluctuations in backscatter values between different time periods. Consequently, smaller, or less pronounced flood signals may fall below the elevated threshold and go undetected, resulting in false negatives. However, decreasing the threshold in flood mapping can lead to an increase in falsely predicted flooded locations, as it becomes more permissive and may interpret noise or minor fluctuations in backscatter values as flood signals. Therefore, a careful balance with a range of 1.01 to 1.05 is struck when setting the threshold to ensure optimal flood detection performance while minimizing false negatives.
2.5 Change detection approach with Otsu method
Change detection using the Otsu method involves the application of histogram thresholding, a well-established technique in digital image processing, to distinguish flooded areas from the rest. This method, introduced by Otsu in 1979, aims to identify an optimal threshold for segmenting the image into distinct categories [107,108]. Flooded regions, characterized by low backscatter values in VV, are recognized by pixels falling below this threshold. Determining the threshold entails the selection of a representative subset within the study area, which should encompass both flooded and non-flooded pixels [109,110]. After identification, the threshold values for each polarization combination are computed based on histogram characteristics. Following the initial preparation of the flood map, final refinement, like those described for the CDA method, was applied to enhance accuracy.
2.6 Supervised machine learning classification
In supervised classification techniques, the user must collect a set of training sample pixels representing each class (flooded and non-flooded). These training samples provide the necessary information about the distribution of pixel reflection values for both flooded and non-flooded locations, enabling the machine learning classifier to classify the entire image [111,112]. In this study, we initially gathered samples of backscatter coefficient values for each land cover class of interest, including water, roads, vegetation, and buildings, with 25 sampled pixel features for each. After gathering training samples for each category, they were combined into a unified feature collection within the GEE platform to streamline classifier training. To train the model, 75 flooded and 100 non-flooded training pixel values have been used. This choice was based on the spatial resolution of Sentinel-1 SAR imagery (10 meters) and the size of the Houston metropolitan area, which includes varied land cover types. A slightly higher number of non-flooded pixels was selected to reflect the more common occurrence of non-flooded areas in urban settings, ensuring balanced training data and preventing model bias. Additionally, preliminary tests indicated that this sample size provided an effective balance between computational efficiency and model accuracy, allowing for robust classification without overfitting. The Random Forest (RF) model, a robust machine learning method extensively utilized in classification and regression tasks, was utilized. [113–116]. The RF algorithm creates numerous decision trees, with each tree offering a class prediction. The model’s prediction is determined by the class that receives the most votes [117–119]. After training the RF classifier, we applied it to classify the entire image, resulting in the classification of flooded areas followed by excluding isolated pixels.
2.7 Accuracy assessment
An accuracy assessment was conducted to provide quantitative evidence of the consistency of the output and to evaluate the performance of the models. This assessment encompassed both the initial and refined flood maps, in which isolated pixels and pixels with slope values greater than 5% were removed. The backscatter coefficient values for flooded regions in the validation dataset were ascertained using High-Water Marks (HWM). These measured points from HWM were overlaid with DFO flood predictions, and DFO polygons that intersect with HWM were selected for evaluation. The validation sample area covered approximately 15 km² within the Houston metropolitan region. These were spatially distributed across areas where High-Water Marks (HWM) intersected with DFO flood polygons (S1 Fig). Model performance was evaluated with the following measures: percent accurately predicted, precision, recall, F1-score and confusion matrix [120–122]. The percentage accurately predicted signifies the proportion of correctly identified flooded areas in relation to the total predictions made by the models. This metric provides a measure of overall model performance, reflecting the ability to accurately discern flood-affected regions from the entire dataset.
Precision, which represents the ratio of correctly predicted positive instances (true positives, Tp) to the total instances predicted as positive (true positives plus false positives, Tp + Fp), serves as an indicator of the model’s accuracy in identifying flooded areas among the positive predictions. A higher precision value implies a reduced likelihood of false positives, thus indicating a more reliable flood detection capability. Recall, often referred to as sensitivity or the true positive rate, measures the ratio of correctly predicted positive instances (true positives, Tp) to the total actual positive instances (true positives plus false negatives, Tp + Fn). Recall assesses the model’s effectiveness in capturing all flood-affected regions within the dataset, highlighting its ability to minimize false negatives and not miss actual flooded locations.
The F1-score is the harmonic mean of precision and recall, providing a balanced assessment of a model’s performance. This accounts for both false positives and false negatives, offering a comprehensive evaluation of how well the model handles true positives while mitigating the impact of misclassifications. A confusion matrix is a pivotal tool for evaluating the performance of classification models. It categorizes predictions into four groups: true positives (Tp), true negatives (Tn), false positives (Fp), and false negatives (Fn) [101,123]. This matrix visually represents the relationships between actual and predicted instances, forming the basis for calculating metrics such as precision, recall, and the F1-score. The distribution of these groups as a percentage underscores the overall prediction accuracy of the models, while precision, recall, and the F1-score delve into the nuances of the models’ ability to predict flooded areas. The confusion matrix was created to provide a visual and quantitative representation of Tp, Tn, Fp, and Fn. This output analysis offers a robust methodology for ensuring a comprehensive evaluation of flood detection models. The reference flood mask for accuracy assessment was developed using the Dartmouth Flood Observatory (DFO) flood map, in combination with high-confidence High-Water Marks (HWMs) collected during Hurricane Harvey. DFO flood polygons intersecting HWM points were selected to ensure that validation was based on actual observed flood extents, enhancing the reliability of the ground truth data.
3 Results and discussion
Table 1 presents a summary of flood mapping model performance for the CDA, Otsu and the ML classification methods. Across the range of threshold values, spanning from 1.01 to 1.05, the table suggests how model accuracy varies for the entire study area as well as for urban and peri-urban regions [52,56]. Of significant interest is the threshold setting of 1.03, where the CDA model demonstrates noteworthy accuracy levels. This is likely due to the optimal distribution of modeled flooded locations for the pixels that show temporal changes in backscatter value above 1.03, with minimal disagreement with the validation data. In contrast, lower thresholds result in fewer identified flooded locations, while higher thresholds yield more. As a result of this, the areas correctly predicted as flooded locations for the CDA and ML methods are 96.11 km² and 81.96 km², accounting for 63.71% and 54.02% accuracy, respectively (Table 1).
For the Otsu method, correctly predicted flooded areas encompass 79.05 km², representing 49.95% accuracy. Specifically, the Otsu method has 55.84% accuracy within urban settings and 61.16% in peri-urban areas. Within the urban area, the CDA achieves an accuracy of 53.82%, while in the peri-urban zone performs with an accuracy of 68.00%. For comparison, the results for the ML model performed better in the urban area, with an accuracy of 58.44% and 65.69% in the peri-urban domain. Reduced accuracy in the models for dense urban areas may stem from the significant heterogeneity in backscatter values, which is caused by the variability in the material composition of the land covers [52,53]. These findings highlight the distinctions in performance between methods with varying landcover, underscoring the importance of tailored approaches to flood detection. Across all cases, it is notable that the model performance is consistently worse in dense urban areas compared to peri-urban areas characterized by more open spaces. While teasing out the exact cause for these trends is challenging due to the complexities of the urban environment, some possible reasons for these trends can be attributed to the increased complexity of dense urban areas, including factors such as intricate infrastructure, building shadows, and high-density features and other built surfaces which can pose challenges for accurate flood detection using satellite imagery [48,50]. This trend may be further exacerbated by high heterogeneity or variability of landcover in urban areas, resulting in highly variable backscatter reflectance in Sentinel-1 images for all model performances. In contrast, peri-urban areas, with their relatively less complex landscape, appear to provide more straightforward conditions for the models to operate effectively.
Fig 3 shows the flood predictions generated by the CDA and ML models for three distinct categories: correctly identified flooded areas, reference flooded locations not identified as flooded by the model, and instances identified as flooded by the model that are not flooded in the reference data for both the CDA (Fig 3a) and ML (Fig 3b). These results reveal a significant number of correctly predicted flooded locations in the open areas in the western part of the study area. In contrast, both models show poorer performance in the central section of the study area, which is dense urban land cover relative to the peri-urban areas in the northeastern, eastern, and southeastern of the city. Comparing the CDA and ML classifiers, the ML model predicts a higher number of flooded locations in most dense urban zones, especially in the central area, but it also produces more false flood predictions. ML classifiers may produce more false flood locations than change detection methods because they incorporate the SAR band signal as input data feature, which can misinterpret complex urban elements for flooding, especially if the training data isn’t sufficiently representative of varied urban scenarios with optimized size and robust source. CDA, relying on temporal changes in pixel values with 63.71% accuracy, typically have fewer parameters to adjust and might be less prone to such errors. Conversely, in peri-urban areas, both models show similar quantities of both correctly and falsely predicted flooded locations. In peri-urban areas, which typically feature fewer complex landscapes than dense urban centers, both ML classifiers and CDA tend to show similar performance in terms of correct and false flood predictions. This can be attributed to the simpler and more uniform land cover which results in less spectral confusion for the models. With fewer buildings and infrastructure, there are fewer sources of error like shadows or varied materials that can lead to misclassification. The relative homogeneity of peri-urban areas means that satellite data in these regions are more straightforward to analyze, leading to both models performing with a similar degree of accuracy.
Model performance through pie chart (c) and histogram (d) representation of the confusion matrix by CDA. For ML classifier pie chart (e) and histogram representation of the confusion matrix (f).
The performance of the CDA (3c and 3d) and ML classifiers (3e and 3f) is represented using pie and bar charts of the confusion matrices. True flooded locations, where the observed and modeled flooded areas match for the CDA, were found to cover 5.53% of the study area, equivalent to 96.11 km². In the case of the ML classifier, the true flooded locations were determined to occupy 4.67% of the study area, totaling 81.16 km², which is 14.95 km² or approximately 1.34% lower than the CDA. The percentage of false non-flooded areas, where the models failed to identify flooded locations, and false flooded areas, where the models wrongly identified flooded areas, were 0.75% and 1.06% for CDA, covering 13.07 km² and 18.45 km², respectively. In the case of ML, these percentages were 1.11% and 1.71%, corresponding to areas of 19.22 km² and 29.73 km², respectively. The CDA demonstrated a lower rate of false non-flooded areas at 0.75% compared to the ML classifier’s 1.11%. However, when it comes to false flooded areas, the CDA had a slightly lower rate at 1.06% compared to the ML classifier’s 1.71%. Additionally, in terms of the areas affected, the CDA covered 13.07 km² for false non-flooded and 18.45 km² for false flooded, while the ML classifier encompassed 19.22 km² for false non-flooded and 29.73 km² for false flooded areas. These differences highlight the varying strengths and weaknesses of the two models in their ability to accurately identify flooded locations.
3.1 Flooded locations at urban and peri-urban areas from CDA
Fig 4 shows flood extents, with 20.25 km² flooded in urban areas and 34.23 km² in peri-urban regions. Correctly predicted floods cover 52.82% of urban areas and 68% of peri-urban areas (Table 1), while false predictions cover 9.75 km² and 6.03 km², respectively. The confusion matrix (Fig 4b–e) shows correct flooded areas covering 3.80% of the study (13.02 km²). False non-flooded and flooded locations cover 2.85% and 1.94% in urban areas, and 2.42% and 4.03% in peri-urban areas, respectively. The CDA model exhibits varying accuracy levels in predicting flooded locations within dense urban and peri-urban land cover types, with higher accuracy observed in the latter. The differences in accuracy between these areas can be attributed to the complexities of urban environments. Dense urban areas often feature intricate structures, shadows, drainage systems, and rapid changes, making it challenging for a simple threshold-based method like CDA to accurately detect floods. Urban areas also tend to exhibit more noise and speckle in radar imagery, which can impact detection performance. The model showcased strong performance in the western, northeastern, and southeastern parts of the study area, which are closer to more rural locations. This effectiveness is likely due to the less heterogeneous land cover in these areas, characterized by more open spaces and a higher presence of non-urban surfaces. These regions exhibit less variation and noise in backscatter signals from the surface, simplifying the task of identifying flooded areas accurately. The reduced complexity in these landscapes compared to urban environments allows for clearer differentiation between water and land, enhancing model accuracy in flood mapping. A visual assessment of the model’s performance in these regions demonstrate its ability to capture flooded areas more accurately in less complex landscapes.
3.2 Flooded locations at urban and peri-urban areas from ML classifier
Fig 5a a shows the extent of flooding in urban (red) and peri-urban (pink) regions based on the ML classifier. The total flooded area is 24.41 km² in urban regions and 36.76 km² in peri-urban zones. Correctly predicted flooded locations account for 58.44% in urban areas and 65.69% in peri-urban areas (Table 1). Actual flooded areas cover 4.07% of the study area—13.93 km² in urban zones and 7.42% (16.25 km²) in peri-urban regions. Falsely predicted flooded regions amount to 9.14 km² (2.67%) in urban areas and 10.71 km² (4.89%) in peri-urban regions. False non-flooded locations cover 2.58% (8.82 km²) for urban and 3.51% (7.69 km²) for peri-urban areas. While the ML classifier demonstrated higher accuracy in urban areas, it also generated more false flooded locations compared to the CDA. The ML model’s improved accuracy in urban areas is mainly due to a higher total count of flooded locations compared to the CDA.
3.3 Comparing CDA and ML method
In urban areas, ML classifier outperformed the CDA in flood detection, though the ML classifier had more false positives compared to CDA. CDA tended to underpredict in urban areas, particularly in the central part of the study area, emphasizing the importance of selecting the appropriate model framework based on land cover types. Comprehensive illustrations of accurately and inaccurately predicted flood sites are presented through confusion matrices in Figs 4 and 5. Both models exhibited enhanced performance in peri-urban areas, achieving accuracy rates of 68.00% and 65.69% in correctly predicting flood locations. The decline in model performance from peri-urban to urban areas can be attributed to the increased prevalence of impervious surfaces and vertical structures, causing amplified double-bounce effects and challenges for accurate flood detection. Additionally, both the CDA and ML models show reduced accuracy near water bodies in the northeast, Center east, and southeast regions. This decreased performance may be due to the unique spectral characteristics and dynamic nature of water bodies, which can complicate flood detection, especially when models are not specifically calibrated for these environments. Model efficacy is significantly influenced by the quality and distribution of validation points, leading to substantial variability in model performance. Therefore, the thoughtful selection and acquisition of robust validation data are pivotal. Performance metrics, including precision, recall, and F1-score, were analyzed. In the entire study area, CDA achieved remarkable precision (approximately 0.837), signifying 83.70% accuracy in identifying flooded areas, with a recall of around 0.877 (87.76%) and an F1-score of about 0.857 (85.70%). For the ML classifier in the same area, precision was about 0.732, recall around 0.809, and the F1-score about 0.768, highlighting its commendable performance in flood detection. In urban areas, CDA showed a precision of 0.572, recall of 0.662, and an F1-score of 0.614, while peri-urban regions displayed a precision of 0.637, recall of 0.720, and an F1-score of 0.676, emphasizing CDA’s versatility in detecting floods. The ML classifier had a precision of 0.604 in urban areas, recall of 0.612, and an F1-score of 0.608. In peri-urban areas, the ML classifier exhibited a precision of 0.625, recall of 0.695, and an F1-score of 0.658, demonstrating its effectiveness in both urban and peri-urban landscapes.
These findings highlight the potential of both approaches for accurate flood detection across various land cover types. The difference in performance between ML classifiers and the CDA in urban versus peri-urban flood detection arises from the ML’s ability to discern complex patterns and the CDA’s threshold-based limitations. Urban areas, with their diverse structures, generate complicated radar signals that ML models can better interpret, leading to superior but less precise results. The simpler landscapes of peri-urban areas allow both models to perform well due to clearer radar signals. The effectiveness of each method varies with the environment’s complexity, stressing the importance of method selection tailored to specific land cover types. Beyond numerical values, these metrics hold profound implications. Precision values indicate the proportion of correctly identified flooded areas among all predicted positive instances, contributing to accurate flood mapping. Recall signifies the proportion of actual flooded areas correctly identified by the model, indicating effective flood detection. The F1-score balances precision and recall, offering a comprehensive measure of a model’s performance. Our analysis showcases the CDA’s consistency and ML classifiers’ balanced performance, contributing to effective flood mapping in diverse landscapes. An insightful understanding of these models’ capabilities is essential for informed decision-making in disaster management. Leveraging the CDA’s reliability and ML classifiers’ balanced performance, decision makers can formulate targeted strategies for flood monitoring, response, and mitigation.
CDA relies on obtaining and processing satellite images, such as SAR images from the GEE platform. The process includes acquiring relevant data, image preprocessing, and setting suitable change detection thresholds. ML Classification, on the other hand, requires data collection for training the model. This involves gathering labeled data, typically with human input, to identify and classify flood and non-flood instances. This process comprises data collection, labeling, and data preparation for model training. The CDA approach generally requires minimal human intervention, as it relies on predefined thresholds and algorithms. In contrast, ML can vary in processing time, from a few hours for small datasets to several days or weeks for larger, more complex datasets. The CDA approach is less complex than ML classification since it doesn’t involve training or fine-tuning a model, whereas ML involves complexity due to the algorithms, feature engineering, and model selection. It demands expertise in ML techniques and potentially more computational resources. The accuracy of the CDA can vary based on the chosen threshold and image quality, but it doesn’t rely on training data or complex algorithms. In comparison, the accuracy of the ML classification approach depends on data quality, algorithm selection, and model performance. It has the potential to achieve higher accuracy, particularly with diverse and accurately labeled training data. Efforts for near-real-time change detection are primarily focused on data acquisition and processing, while non-real-time ML classification involves additional steps such as data collection, labeling, training, and model development. The CDA can provide quick results but may have lower accuracy. Although the model’s performance in dense urban areas is limited, it still offers valuable insights into the uncertainty and challenges associated with flood detection in complex environments. This understanding can be crucial for emergency responders and decision-makers as part of broader flood risk assessments. The model provides a foundation for capturing variability, particularly in peri-urban areas, but further refinement is necessary to enhance its reliability in urban settings. Future work should focus on improving accuracy and reducing uncertainty through advanced techniques and better data integration.
Although the model produces false positives near water bodies, understanding these uncertainties is a critical step toward improving information for flood managers and advancing detection strategies. By highlighting areas where the model’s performance needs refinement, particularly in urban environments, flood detection can be better understood and advanced further, ultimately enabling extracting critical information from remotely sensed data for emergency responders and planners can use to guide future adjustments. While we are not yet at a stage where the model can be fully deployed for real-time response, this work is a significant step forward by identifies key areas for further development. In peri-urban areas, where the model performs more effectively, these findings have the potential to contribute to ongoing efforts to enhance resilience and flood risk management.
Unlike previous research, our study encompasses the full extent of the Houston area, utilizing the large volume of validation data and providing detailed insights into both dense urban and peri-urban regions. For example, Banolia et. al. (2023) [49], Hamidi et.al. (2023) [33] and Mehravar et. al. (2023) [37] studied flooding and its impact in urban areas, however, did not extensively analysis the strengths and weaknesses of the models in determining the flood locations in the densely urban and peri-urban areas. This broader scope allows for a more complete understanding of flood dynamics across varying land cover types. In terms of model performance, our study not only addresses the entire study area but also specifically evaluates the performance in dense urban and peri-urban areas separately. This nuanced approach allows for a more targeted analysis and understanding of the models’ effectiveness in different environments. Overall, compared to previous literature, our study offers a more extensive and detailed examination of flood detection in urban and peri-urban settings. The depth of our accuracy assessment and the scale of our study area set our research apart, providing valuable insights for effective flood management and disaster response strategies in these critical regions.
4 Conclusion
This study evaluated three flood detection methodologies CDA, Otsu method, and ML classification across urban and peri-urban landscapes. The results highlight that CDA provides reliable performance, particularly in peri-urban areas, due to the simplicity of land cover. While CDA performed adequately in urban areas, ML classifiers showed higher accuracy in complex urban environments, though they also generated more false positives. The Otsu method demonstrated lower overall accuracy in both urban and peri-urban regions. CDA is an efficient option for quick flood detection with minimal data and resources, while ML classifiers offer higher accuracy but require more extensive data and processing. The choice of methodology should be guided by specific requirements and the resources available for flood management.
Despite these findings, several limitations were noted, including the potential for false positives, the challenges of SAR imagery in urban environments, and the dependency on validation data quality. Future work should focus on improving model validation through enhanced data integration, such as combining SAR with hydrological data, and exploring ensemble modeling techniques to increase flood detection accuracy. Although the model’s performance in dense urban areas is limited, it still offers valuable insights into the uncertainty and challenges associated with flood detection in complex environments. This understanding can be crucial for emergency responders and decision-makers as part of broader flood risk assessments. The model provides a foundation for capturing variability, particularly in peri-urban areas, but further refinement is necessary to enhance its reliability in urban settings. Future work should focus on improving accuracy and reducing uncertainty through advanced techniques and better data integration. This study is based on a single flood event—Hurricane Harvey in Houston—which provides a detailed case for evaluating method performance. However, the findings may not fully generalize across flood types or geographic regions. Future work should extend the analysis to include multiple flood events with varying hydrological and land cover characteristics to validate and enhance the robustness of these approaches.
Supporting information
S1 Fig. Spatial distribution of flooded and non-flooded areas of the reference map.
The base map layer was obtained from ArcGIS Basemap, and the administrative boundary shapefile was sourced from the City of Houston GIS (COHGIS) Open Data Portal (COH Administrative Boundaries dataset, https://cohgismycity.opendata.arcgis.com/).
https://doi.org/10.1371/journal.pwat.0000269.s001
(TIFF)
S2 Fig. Backscatter ratio histogram for threshold selection in change detection approach.
https://doi.org/10.1371/journal.pwat.0000269.s002
(TIFF)
References
- 1. Dutta D, Herath S, Musiake K. A mathematical model for flood loss estimation. J Hydrol. 2003;277(1):24–49.
- 2. Hallegatte S, Green C, Nicholls RJ, Corfee-Morlot J. Future flood losses in major coastal cities. Nat Clim Change. 2013;3(9):802–6.
- 3. Karamouz M, Fereshtehpour M, Ahmadvand F, Zahmatkesh Z. Coastal flood damage estimator: An alternative to FEMA’s HAZUS platform. J Irrig Drain Eng. 2016;142(6):04016016.
- 4. Li C, Cheng X, Li N, Du X, Yu Q, Kan G. A framework for flood risk analysis and benefit assessment of flood control measures in urban areas. Int J Environ Res Public Health. 2016;13(8):787. pmid:27527202
- 5. Rainey JL, Brody SD, Galloway GE, Highfield WE. Assessment of the growing threat of urban flooding: a case study of a national survey. Urban Water J. 2021;18(5):375–81.
- 6. Wang L, Huang Z, Gan B, Zhang Z, Fu H, Fang D, et al. Climate change impacts on magnitude and frequency of urban floods under scenario and model uncertainties. J Environ Manage. 2024;366:121679. pmid:38996601
- 7.
Jha AK, Bloch R, Lamond J. Cities and flooding: a guide to integrated urban flood risk management for the 21st century. World Bank Publications; 2012.
- 8.
Holden E. Flooding will affect double the number of people worldwide by 2030. The Guardian. 2020 [Accessed 2023 August 18]. https://www.theguardian.com/environment/2020/apr/23/flooding-double-number-people-worldwide-2030
- 9. Neumann B, Vafeidis AT, Zimmermann J, Nicholls RJ. Future coastal population growth and exposure to sea-level rise and coastal flooding--a global assessment. PLoS One. 2015;10(3):e0118571. pmid:25760037
- 10. Alfieri L, Bisselink B, Dottori F, Naumann G, de Roo A, Salamon P, et al. Global projections of river flood risk in a warmer world. Earth’s Future. 2017;5(2):171–82.
- 11. Kumar SV, Harrison KW, Peters-Lidard CD, Santanello JA, Kirschbaum D. Assessing the impact of L-band observations on drought and flood risk estimation: A decision-theoretic approach in an OSSE environment. J Hydrometeorol. 2014;15(6):2140–56.
- 12. Sy M a i H, Luu C, Bui QD, Ha H, Nguyen DQ. Urban flood risk assessment using Sentinel-1 on the google earth engine: a case study in Thai Nguyen city, Vietnam. Remote Sens Appl Soc Environ. 2023;31:100987.
- 13. Grimaldi S, Li Y, Pauwels VRN, Walker JP. Remote sensing-derived water extent and level to constrain hydraulic flood forecasting models: opportunities and challenges. Surv Geophys. 2016;37(5):977–1034.
- 14. Kansakar P, Hossain F. A review of applications of satellite earth observation data for global societal benefit and stewardship of planet earth. Space Policy. 2016;36:46–54.
- 15. Papageorgiou E, Foumelis M, Mouratidis A. Earth observation data synergy for the enhanced monitoring of ephemeral water bodies to anticipate karst-related flooding. GeoHazards. 2023;4(2):197–216.
- 16. Rahat SH, Steissberg T, Chang W, Chen X, Mandavya G, Tracy J, et al. Remote sensing-enabled machine learning for river water quality modeling under multidimensional uncertainty. Sci Total Environ. 2023;898:165504. pmid:37459982
- 17. Colosio P, Tedesco M, Tellman E. Flood monitoring using enhanced resolution passive microwave data: a test case over Bangladesh. Remote Sens. 2022;14(5):1180.
- 18. Konapala G, Kumar SV, Khalique Ahmad S. Exploring Sentinel-1 and Sentinel-2 diversity for flood inundation mapping using deep learning. ISPRS J Photogramm Remote Sens. 2021;180:163–73.
- 19. Samela C, Coluzzi R, Imbrenda V, Manfreda S, Lanfredi M. Satellite flood detection integrating hydrogeomorphic and spectral indices. GIScience Remote Sens. 2022;59(1):1997–2018.
- 20. Ticehurst C, Guerschman JP, Chen Y. The strengths and limitations in using the daily MODIS open water likelihood algorithm for identifying flood events. Remote Sens. 2014;6(12).
- 21. Farhadi H, Kiani A, Ebadi H, Asgary A. Development of an automatic time-series flood mapping framework using Sentinel-1 and 2 imagery. Stoch Environ Res Risk Assess. 2025;39(6):2627–55.
- 22. Dube T, Seaton D, Shoko C, Mbow C. Advancements in earth observation for water resources monitoring and management in Africa: a comprehensive review. J Hydrol. 2023;623:129738.
- 23. Thomas M, Tellman E, Osgood DE, DeVries B, Islam AS, Steckler MS. A framework to assess remote sensing algorithms for satellite-based flood index insurance. IEEE J Sel Top Appl Earth Obs Remote Sens. 2023;16:2589–604.
- 24. Berz G, Kron W, Loster T, Rauch E, Schimetschek J, Schmieder J, et al. World map of natural hazards – A global view of the distribution and intensity of significant exposures. Natural Hazards. 2001;23(2–3):443–65.
- 25. Munawar HS, Hammad AWA, Waller ST. Remote sensing methods for flood prediction: a review. Sensors (Basel). 2022;22(3):960. pmid:35161706
- 26. Zhou Y, Dong J, Xiao X, Xiao T, Yang Z, Zhao G. Open surface water mapping algorithms: a comparison of water-related spectral indices and sensors. Water. 2017;9(4):256.
- 27. Albertini C, Gioia A, Iacobellis V, Manfreda S. Detection of surface water and floods with multispectral satellites. Remote Sens. 2022;14(23):6005.
- 28. Rad AM, Kreitler J, Sadegh M. Augmented normalized difference water index for improved surface water monitoring. Environ Model Softw. 2021;140:105030.
- 29. Farhadi H, Ebadi H, Kiani A, Asgary A. Introducing a new index for flood mapping using Sentinel-2 imagery (SFMI). Comput Geosci. 2025;194:105742.
- 30. Farhadi H, Ebadi H, Kiani A, Asgary A. A novel flood/water extraction index (FWEI) for identifying water and flooded areas using sentinel-2 visible and near-infrared spectral bands. Stoch Environ Res Risk Assess. 2024;38:1–23.
- 31. Qi L, Wang M, Hu C, Holt B. On the capacity of Sentinel-1 synthetic aperture radar in detecting floating macroalgae and other floating matters. Remote Sens Environ. 2022;280:113188.
- 32. Chini M, Papastergios A, Pulvirenti L, Pierdicca N, Matgen P, Parcharidis I. SAR coherence and polarimetric information for improving flood mapping. In: 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS). 2016: 7577–80.
- 33. Hamidi E, Peter BG, Muñoz DF, Moftakhari H, Moradkhani H. Fast flood extent monitoring with SAR change detection using Google Earth Engine. IEEE Transact Geosci Remote Sens. 2023;61:1–19.
- 34. Mason DC, Bevington J, Dance SL, Revilla-Romero B, Smith R, Vetra-Carvalho S. Improving urban flood mapping by merging synthetic aperture radar-derived flood footprints with flood hazard maps. Water. 2021;13(11):1577.
- 35. Li Z, Demir I. U-net-based semantic classification for flood extent extraction using SAR imagery and GEE platform: a case study for 2019 central US flooding. Sci Total Environ. 2023;869:161757. pmid:36690091
- 36. Manavalan R. SAR image analysis techniques for flood area mapping - literature survey. Earth Sci Inform. 2016;10(1):1–14.
- 37. Mehravar S, Razavi-Termeh SV, Moghimi A, Ranjgar B, Foroughnia F, Amani M. Flood susceptibility mapping using multi-temporal SAR imagery and novel integration of nature-inspired algorithms into support vector regression. J Hydrol. 2023;617:129100.
- 38.
Flood Monitoring from SAR Data | SpringerLink [Internet]. [Accessed 2024 April 8]. https://doi.org/10.1007/978-90-481-9618-0_3
- 39. Berezowski T, Bieliński T, Osowicki J. Flooding extent mapping for synthetic aperture radar time series using river gauge observations. IEEE J Sel Top Appl Earth Obs Remote Sens. 2020;13:2626–38.
- 40. Moity N, Delgado B, Salinas-de-León P. Mangroves in the Galapagos Islands: distribution and dynamics. PLoS One. 2019;14(1):e0209313. pmid:30625180
- 41. Alexakis DD, Stavroulaki EG, Tsanis IK. Using Sentinel-1A DInSAR interferometry and Landsat 8 data for monitoring water level changes in two lakes in Crete, Greece. Geocarto Int. 2018;34(7):703–21.
- 42. Google Earth Engine for large-scale flood mapping using SAR data and impact assessment on agriculture and population of Ganga-Brahmaputra basin. Sustainability. 14(7):4210.
- 43. Hostache R, Chini M, Giustarini L, Neal J, Kavetski D, Wood M, et al. Near‐Real‐time assimilation of SAR‐derived flood maps for improving flood forecasts. Water Res Res. 2018;54(8):5516–35.
- 44. DeVries B, Huang C, Armston J, Huang W, Jones JW, Lang MW. Rapid and robust monitoring of flood events using Sentinel-1 and Landsat data on the Google Earth Engine. Remote Sens Environ. 2020;240:111664.
- 45. Tiwari V, Kumar V, Matin MA, Thapa A, Ellenburg WL, Gupta N, et al. Flood inundation mapping- Kerala 2018; Harnessing the power of SAR, automatic threshold detection method and Google Earth Engine. PLoS One. 2020;15(8):e0237324. pmid:32813701
- 46. Shen X, Wang D, Mao K, Anagnostou E, Hong Y. Inundation extent mapping by synthetic aperture radar: a review. Remote Sens. 2019;11(7):879.
- 47. Mason DC, Dance SL, Cloke HL. Floodwater detection in urban areas using Sentinel-1 and WorldDEM data. J Appl Remote Sens. 2021;15(3):032003.
- 48. Li Y, Martinis S, Wieland M. Urban flood mapping with an active self-learning convolutional neural network based on TerraSAR-X intensity and interferometric coherence. ISPRS J Photogramm Remote Sens. 2019;152:178–91.
- 49.
Banolia C, Prabhakar KR, Deshpande S. Monitoring urban flooding using SAR—A Mumbai case study. In: Sharma N, Goje A, Chakrabarti A, Bruckstein AM, eds. Data management, analytics and innovation. Singapore: Springer Nature; 2023: 59–68.
- 50. Schumann GJP, Neal JC, Mason DC, Bates PD. The accuracy of sequential aerial photography and SAR data for observing urban flood dynamics, a case study of the UK summer 2007 floods. Remote Sens Environ. 2011;115(10):2536–46.
- 51. Boemke B, Turki I, Brüll C, Lehmkuhl F. Assessing complex aeolian dune field morphology and evolution with Sentinel-1 SAR imagery – possibilities and limitations. Aeolian Res. 2023;62:100876.
- 52. Vanama KR, Rao YS. Change detection based flood mapping of 2015 flood event of Chennai city using Sentinel-1 SAR images. In: IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium. 2019: 9729–32.
- 53. Semenzato A, Pappalardo SE, Codato D, Trivelloni U, De Zorzi S, Ferrari S. Mapping and monitoring urban environment through Sentinel-1 SAR data: a case study in the Veneto region (Italy). ISPRS Int J Geo Inf. 2020;9(6):375.
- 54. Malmgren-Hansen D, Sohnesen T, Fisker P, Baez J. Sentinel-1 change detection analysis for cyclone damage assessment in urban environments. Remote Sens. 2020;12(15):2409.
- 55. Singha M, Dong J, Zhang G, Xiao X. High resolution paddy rice maps in cloud-prone Bangladesh and Northeast India using Sentinel-1 data. Sci Data. 2019;6(1):26.
- 56. Agnihotri AK, Ohri A, Gaur S, Shivam , Das N, Mishra S. Flood inundation mapping and monitoring using SAR data and its impact on Ramganga River in Ganga basin. Environ Monit Assess. 2019;191(12):760. pmid:31745827
- 57. Koukiou G. SAR features and techniques for urban planning—A review. Remote Sens. 2024;16(11):1923.
- 58. Li Y, Martinis S, Wieland M, Schlaffer S, Natsuaki R. Urban flood mapping using SAR intensity and interferometric coherence via Bayesian network fusion. Remote Sens. 2019;11(19):2231.
- 59. Garg S, Dasgupta A, Motagh M, Martinis S, Selvakumaran S. Unlocking the full potential of Sentinel-1 for flood detection in arid regions. Remote Sens Environ. 2024;315:114417.
- 60. Horritt MS, Mason DC, Cobby DM, Davenport IJ, Bates PD. Waterline mapping in flooded vegetation from airborne SAR imagery. Remote Sens Environ. 2003;85(3):271–81.
- 61. Huang L, Li Z, Tian BS, Chen Q, Liu JL, Zhang R. Classification and snow line detection for glacial areas using the polarimetric SAR image. Remote Sens Environ. 2011;115(7):1721–32.
- 62. van der Wal D, Herman PMJ, Wielemaker-van den Dool A. Characterisation of surface roughness and sediment texture of intertidal flats using ERS SAR imagery. Remote Sens Environ. 2005;98(1):96–109.
- 63. Bovolo F, Bruzzone L. The time variable in data fusion: a change detection perspective. IEEE Geosci Remote Sens Mag. 2015;3(3):8–26.
- 64. Dellepiane SG, Angiati E. A new method for cross-normalization and multitemporal visualization of SAR images for the detection of flooded areas. IEEE Transact Geosci Remote Sens. 2012;50(7):2765–79.
- 65. Melkamu T, Bagyaraj M, Adimaw M, Ngusie A, Karuppannan S. Detecting and mapping flood inundation areas in Fogera-Dera floodplain, Ethiopia during an extreme wet season using Sentinel-1 data. Phys Chem Earth Parts ABC. 2022;127:103189.
- 66. Slesinski J, Wierzbicki D, Kedzierski M. Application of multitemporal change detection in radar satellite imagery using REACTIV-based method for geospatial intelligence. Sensors (Basel). 2023;23(10):4922. pmid:37430836
- 67. Lê TT, Froger JL, Ho Tong Minh D. Multiscale framework for rapid change analysis from SAR image time series: Case study of flood monitoring in the central coast regions of Vietnam. Remote Sens Environ. 2022;269:112837.
- 68. Motta M, de Castro Neto M, Sarmento P. A mixed approach for urban flood prediction using machine learning and GIS. Int J Disaster Risk Reduct. 2021;56:102154.
- 69. Norallahi M, Seyed Kaboli H. Urban flood hazard mapping using machine learning models: GARP, RF, MaxEnt and NB. Nat Hazards. 2021;106(1):119–37.
- 70. Marzialetti F, Gamba P, Sorriso A, Carranza ML. Monitoring urban expansion by coupling multi-temporal active remote sensing and landscape analysis: changes in the metropolitan area of Cordoba (Argentina) from 2010 to 2021. Remote Sens. 2023;15(2):336.
- 71. Sahana M, Ravetz J, Patel PP, Dadashpoor H, Follmann A. Where is the peri-urban? A systematic review of peri-urban research and approaches for its identification and demarcation worldwide. Remote Sens. 2023;15(5):1316.
- 72. Sun G, Cheng J, Zhang A, Jia X, Yao Y, Jiao Z. Hierarchical fusion of optical and dual-polarized SAR on impervious surface mapping at city scale. ISPRS J Photogramm Remote Sens. 2022;184:264–78.
- 73. Miller JD, Kim H, Kjeldsen TR, Packman J, Grebby S, Dearden R. Assessing the impact of urbanization on storm runoff in a peri-urban catchment using historical change in impervious cover. J Hydrol. 2014;515:59–70.
- 74. Reinstaller S, Krebs G, Pichler M, Muschalla D. Identification of high-impact uncertainty sources for urban flood models in hillside peri-urban catchments. Water. 2022;14(12):1973.
- 75. Stefanidis S, Alexandridis V, Mallinis G. A cloud-based mapping approach for assessing spatiotemporal changes in erosion dynamics due to biotic and abiotic disturbances in a Mediterranean peri-urban forest. CATENA. 2022;218:106564.
- 76. Chung LC, Xie J, Ren C. Improved machine-learning mapping of local climate zones in metropolitan areas using composite Earth observation data in Google Earth Engine. Build Environ. 2021;199:107879.
- 77. Pourghasemi HR, Amiri M, Edalat M, Ahrari AH, Panahi M, Sadhasivam N. Assessment of urban infrastructures exposed to flood using susceptibility map and Google Earth Engine. IEEE J Sel Top Appl Earth Obs Remote Sens. 2021;14:1923–37.
- 78. Saki SA, Sofia G, Anagnostou EN. Characterizing CONUS-wide spatio-temporal changes in daily precipitation, flow, and variability of extremes. J Hydrol. 2023;626:130336.
- 79. Mayer T, Poortinga A, Bhandari B, Nicolau AP, Markert K, Thwal NS, et al. Deep learning approach for Sentinel-1 surface water mapping leveraging Google Earth Engine. ISPRS Open J Photogramm Remote Sens. 2021;2:100005.
- 80. Mullissa A, Vollrath A, Odongo-Braun C, Slagter B, Balling J, Gou Y. Sentinel-1 SAR backscatter analysis ready data preparation in Google Earth Engine. Remote Sens. 2021;13(10):1954.
- 81. Nghia BPQ, Pal I, Chollacoop N, Mukhopadhyay A. Applying Google earth engine for flood mapping and monitoring in the downstream provinces of Mekong river. Prog Disaster Sci. 2022;14:100235.
- 82. Schiller AR. The impact of a storm surge on business establishments in the Houston MSA. Nat Hazards. 2010;56(1):331–46.
- 83. Bodenreider C, Wright L, Barr O, Xu K, Wilson S. Assessment of social, economic, and geographic vulnerability pre- and post-hurricane harvey in Houston, Texas. Environ Justice. 2019;12(4):182–93.
- 84. Griego AL, Flores AB, Collins TW, Grineski SE. Social vulnerability, disaster assistance, and recovery: a population-based study of Hurricane Harvey in Greater Houston, Texas. Int J Disaster Risk Reduct. 2020;51:101766.
- 85. Chen M, Li Z, Gao S. Multisensor remote sensing and the multidimensional modeling of extreme flood events. geophysical monograph series. Wiley; 2022: 87–104.
- 86. Valle-Levinson A, Olabarrieta M, Heilman L. Compound flooding in Houston-Galveston Bay during Hurricane Harvey. Sci Total Environ. 2020;747:141272. pmid:32777508
- 87. Zou L, Liao D, Lam NSN, Meyer MA, Gharaibeh NG, Cai H, et al. Social media for emergency rescue: An analysis of rescue requests on Twitter during Hurricane Harvey. Int J Disaster Risk Reduct. 2023;85:103513.
- 88. Zhang W, Villarini G, Vecchi GA, Smith JA. Urbanization exacerbated the rainfall and flooding caused by hurricane Harvey in Houston. Nature. 2018;563(7731):384–8.
- 89. Chini M, Pelich R, Pulvirenti L, Pierdicca N, Hostache R, Matgen P. Sentinel-1 InSAR coherence to detect floodwater in urban areas: Houston and Hurricane Harvey as a test case. Remote Sens. 2019;11(2):107.
- 90.
Tropical Cyclone Report: Hurricane Harvey. [Accessed 2025 April 28]. https://www.nhc.noaa.gov/data/tcr/AL092017_Harvey.pdf
- 91.
Amadeo K. Hurricane Harvey facts, damage and costs.
- 92.
Sentinel-1 - Missions - Sentinel Online. [Accessed 2023 October 12]. https://copernicus.eu/missions/sentinel-1
- 93.
Wickel BA, Lehner B, Sindorf N. HydroSHEDS: a global comprehensive hydrographic dataset. 2007;2007:H11H–05.
- 94. Pekel JF, Cottam A, Gorelick N, Belward AS. High-resolution mapping of global surface water and its long-term changes. Nat. 2016;540(7633):418–22.
- 95.
High-water marks and flooding. U.S. Geological survey. [Accessed 2023 October 2]. https://www.usgs.gov/special-topics/water-science-school/science/high-water-marks-and-flooding
- 96.
Dartmouth Flood Observatory. Dartmouth flood observatory. [Accessed 2023 July 9]. https://floodobservatory.colorado.edu/Archives/index.html
- 97. Lehmler S, Förster M, Frick A. Modelling green volume using Sentinel-1, -2, PALSAR-2 satellite data and machine learning for urban and semi-urban areas in Germany. Environ Manage. 2023;72(3):657–70. pmid:37233749
- 98. Liu J, Li Y, Zhang Y, Liu X. Large-scale impervious surface area mapping and pattern evolution of the yellow river delta using sentinel-1/2 on the gee. Remote Sens. 2023;15(1):136.
- 99.
Google for Developers. NLCD 2019: USGS National Land Cover Database, 2019 release | Earth Engine Data Catalog. [Accessed 2023 July 9]. https://developers.google.com/earth-engine/datasets/catalog/USGS_NLCD_RELEASES_2019_REL_NLCD
- 100. Amani M, Ghorbanian A, Ahmadi SA, Kakooei M, Moghimi A, Mirmazloumi SM, et al. Google earth engine cloud computing platform for remote sensing big data applications: a comprehensive review. IEEE J Sel Top Appl Earth Obs Remote Sens. 2020;13:5326–50.
- 101. Moharrami M, Javanbakht M, Attarchi S. Automatic flood detection using sentinel-1 images on the google earth engine. Environ Monit Assess. 2021;193(5):248. pmid:33825990
- 102. Lee JS, Jurkevich L, Dewaele P, Wambacq P, Oosterlinck A. Speckle filtering of synthetic aperture radar images: a review. Remote Sens Rev. 1994;8(4):313–40.
- 103.
Step-by-step: Recommended practice: Flood mapping and damage assessment using Sentinel-1 SAR data in Google Earth Engine. [Accessed 2023 October 13]. https://www.un-spider.org/advisory-support/recommended-practices/recommended-practice-google-earth-engine-flood-mapping/step-by-step
- 104.
JRC global surface water mapping layers, v1.4 | Earth Engine Data Catalog. [Accessed 2023 July 9]. https://developers.google.com/earth-engine/datasets/catalog/JRC_GSW1_4_GlobalSurfaceWater
- 105. Vanama KR, Rao YS. Change detection based flood mapping of 2015 flood event of Chennai city using Sentinel-1 SAR images. In: IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium. 2019: 9729–32. https://ieeexplore.ieee.org/abstract/document/8899282?casa_token=6zj1KLBkEyAAAAAA:oKWgwb_wYxQpd2t1_8GdTdfcvS1YwDmq4jbSEVlgFBGzUbGR2GwaLAWYT0JLnUMQ0pY-vYTT-w
- 106. Li Y, Martinis S, Plank S, Ludwig R. An automatic change detection approach for rapid flood mapping in Sentinel-1 SAR data. Int J Appl Earth Obs Geoinform. 2018;73:123–35.
- 107. Mudi S, A JP, Das P. Flood hazard mapping in Assam using Sentinel-1 SAR data. In: Shit PK, Pourghasemi HR, Bhunia GS, Das P, Narsimha A, eds. Geospatial technology for environmental hazards: modeling and management in Asian Countries. Cham: Springer International Publishing; 2021: 459–73.
- 108. Tran KH, Menenti M, Jia L. Surface water mapping and flood monitoring in the Mekong delta using Sentinel-1 SAR time series and Otsu threshold. Remote Sens. 2022;14(22):5721.
- 109. Schlaffer S, Matgen P, Hollaus M, Wagner W. Flood detection from multi-temporal SAR data using harmonic analysis and change detection. Int J Appl Earth Obs Geoinform. 2015;38:15–24.
- 110. Tupas ME, Roth F, Bauer-Marschallinger B, Wagner W. An intercomparison of Sentinel-1 based change detection algorithms for flood mapping. Remote Sens. 2023;15(5):1200.
- 111. Bayik C, Abdikan S, Ozbulak G, Alasag T, Aydemir S, Balik Sanli F. Exploiting multi-temporal sentinel-1 sar data for flood extend mapping. Int Arch Photogramm Remote Sens Spat Inf Sci. 2018;XLII(3-W4):109–13.
- 112. Sharifi A. Development of a method for flood detection based on Sentinel‐1 images and classifier algorithms. Water Environ J. 2021;35(3):924–9.
- 113. Bessinger M, Lück-Vogel M, Skowno A, Conrad F. Landsat-8 based coastal ecosystem mapping in South Africa using random forest classification in Google Earth Engine. South Afr J Bot. 2022;150:928–39.
- 114. Oliphant AJ, Thenkabail PS, Teluguntla P, Xiong J, Gumma MK, Congalton RG. Mapping cropland extent of Southeast and Northeast Asia using multi-year time-series Landsat 30-m data using a random forest classifier on the Google Earth Engine Cloud. Int J Appl Earth Obs Geoinform. 2019;81:110–24.
- 115. Singh S, Kansal ML. Chamoli flash-flood mapping and evaluation with a supervised classifier and NDWI thresholding using Sentinel-2 optical data in Google earth engine. Earth Sci Inform. 2022;15(2):1073–86.
- 116. Sunny MR, Kabir MA, Naheen IT, Ahad MT. Residential energy management: A machine learning perspective. In: 2020 IEEE Green Technologies Conference(GreenTech). 2020: 229–34. https://ieeexplore.ieee.org/document/9289737
- 117. Belgiu M, Drăguţ L. Random forest in remote sensing: a review of applications and future directions. ISPRS J Photogramm Remote Sens. 2016;114:24–31.
- 118. Pal M. Random forest classifier for remote sensing classification. Int J Remote Sens. 2005;26(1):217–22.
- 119.
Detecting SARS-CoV-2 From Chest X-Ray Using Artificial Intelligence | IEEE Journals & Magazine | IEEE Xplore. [Accessed 2024 March 12]. https://ieeexplore.ieee.org/document/9360740
- 120. Ghanim AAJ, Shaf A, Ali T, Zafar M, Al-Areeq AM, Alyami SH. An improved flood susceptibility assessment in Jeddah, Saudi Arabia, using advanced machine learning techniques. Water. 2023;15(14):2511.
- 121. Kim D, Park J, Han H, Lee H, Kim HS, Kim S. Application of AI-based models for flood water level forecasting and flood risk classification. KSCE J Civil Eng. 2023;27(7):3163–74.
- 122. Mehedi MAA, Smith V, Hosseiny H, Jiao X. Unraveling the complexities of urban fluvial flood hydraulics through AI. Sci Rep. 2022;12(1):18738. pmid:36333429
- 123. Al-Abadi AM. Mapping flood susceptibility in an arid region of southern Iraq using ensemble machine learning classifiers: a comparative study. Arab J Geosci. 2018;11(9).