Weihui City is located in the northern part of Henan Province, China, covering a total area of 862 km², with a north-south length of 43 km and an east-west width of 35 km (Fig. 1). Situated within the Hai River Basin, the city lies at the confluence of multiple rivers, including the Wei River and the Communism Canal, resulting in a large surface water area. The urban area is built on a plain formed by river sedimentation, featuring low-lying terrain that slopes from east to west, which hinders effective drainage of accumulated water. In recent years, rapid urbanization has altered the characteristics of the underlying surface, further increasing the risk of urban waterlogging. Climatically, Weihui experiences a temperate continental monsoon climate, characterized by hot and rainy summers. From July to August, precipitation accounts for approximately 50 % of the annual total, often concentrated in a few intense rainstorms. Moreover, interannual variation in precipitation is significant – wet years may see rainfall as high as 1000 mm, while dry years may receive only 400 mm, a difference of more than 2.5 times. This results in highly variable river flows within the city, making summer the peak season for flood disasters in Weihui (Du et al., 2024; Zhang et al., 2022).

Starting on 17 July 2021, persistent heavy rainfall triggered widespread flooding across many parts of Henan Province. Records from the Xinxiang Meteorological Station indicated that this extraordinary rainstorm was the most intense since observations began, with a maximum recorded rainfall of 965.5 mm. Located downstream, Weihui City faced enormous pressure from upstream flood discharge even after the rainfall had ceased. River water levels continued to rise, eventually surpassing the warning thresholds. Overflows from the Wei River and the Communism Canal caused massive volumes of floodwater to inundate the urban area of Weihui, leading to extremely severe urban flooding. At its peak, 79 % of the city's building area was submerged, with floodwater volume reaching 21 000 000 m³. All four reservoirs in the city overflowed, making it impossible to discharge floodwaters via the river system. Compounding the issue, the city's power supply was disrupted by the flood, preventing pumping stations from operating at full capacity. As a result, floodwaters in the main urban area were not fully drained until 31 July. Nearly 300,000 people were affected, and over 14 000 houses were inundated. In addition to the urban devastation, all 13 townships under Weihui's jurisdiction experienced varying degrees of flooding, with approximately 330 km² of crops damaged (Zhang et al., 2022).

The Sentinel-1 satellite constellation currently consists of three satellites: Sentinel-1A, launched in 2014; Sentinel-1B, launched in 2016; and Sentinel-1C, launched in 2024. Sentinel-1B was officially decommissioned in August 2022 due to an irreparable power failure (Torres et al., 2012; Geudtner et al., 2014). When operating with a single satellite, Sentinel-1 has a revisit cycle of 12 d, with observation frequency reaching up to 6 d in certain regions (Torres et al., 2012). To meet diverse application needs, Sentinel-1 is equipped with four imaging modes: stripmap mode (SM), interferometric wide swath mode (IW), extra wide swath mode (EW), and wave mode (WV). Among these, the IW mode is the primary mode for land surface observation. It enables coverage of the vast majority of the Earth's land areas, with each acquisition providing a swath width of 250 km and a spatial resolution of 5 × 20 m.

This study primarily utilizes the intensity data from Sentinel-1 ground range detected (GRD) products for flood extraction (Nhangumbe et al., 2023; Torres et al., 2012). The data processing involves the following key steps: (1) Precise orbit correction – applying refined satellite orbit data to improve geolocation accuracy. (2) Thermal noise removal – eliminating sensor-induced noise to enhance image quality. (3) Radiometric calibration (Sigma0) – converting pixel values into physically meaningful backscatter intensity using the sensor calibration metadata embedded in the images. (4) Speckle filtering – applying Refined Lee filtering to reduce speckle noise while preserving image details. (5) Terrain correction – using a digital elevation model (DEM) to correct the influence of surface topography on backscatter intensity, resulting in a normalized backscatter coefficient. (6) Conversion to decibel scale (dB) – expressing the backscatter coefficient on a logarithmic scale to facilitate easier computation and statistical analysis, especially for values that span several orders of magnitude.

The Sentinel-1 images used in this study correspond to Path 40 (112) and Frame 113 (111), covering the period from 4 January to 30 December 2021. The flood event occurred in mid-to-late July, with large-scale floodwaters receding approximately by mid-to-late August.

The Global Precipitation Measurement (GPM) mission, developed through international collaboration among multiple space agencies, enables frequent and high-accuracy global precipitation observations. The satellite is equipped with two main instruments for precipitation monitoring: the dual-frequency precipitation radar (DPR) and the GPM Microwave Imager (GMI). The Integrated Multi-satellite Retrievals for GPM (IMERG) precipitation product, provided by the National Aeronautics and Space Administration (NASA), is one of the key datasets derived from GPM observations (Tang et al., 2020). This product is primarily generated using GPM satellite data and is supplemented with data from other meteorological satellites, including TRMM, METOP, and NOAA (Kummerow et al., 2000). By integrating observations across various frequency bands and resolutions, the IMERG product enhances both the accuracy and spatiotemporal resolution of precipitation estimates. It offers global precipitation data at a temporal resolution of 30 minutes and a spatial resolution of 0.1°. In this study, the “precipitationCal” band from the IMERG product is used as the reference source for precipitation data. This band has been calibrated using ground-based observation station data and is the officially recommended product in the IMERG user manual. GPM data are mainly used to characterize rainfall intensity during the flood event and to assist in validating the flood extraction results.

The Global Surface Water (GSW) dataset is a 30 m resolution product that provides global surface water distribution information spanning the past 38 years (Pekel et al., 2016). It is developed and continuously updated by the Joint Research Center (JRC). Each pixel is classified as either water or non-water using an expert system, and then temporal analysis is applied to capture the historical dynamics of surface water presence. The algorithm proposed in this study uses the occurrence band from the GSW dataset, which represents the frequency of water occurrence for each pixel, with values ranging from 0 % to 100 %. Based on threshold values, surface water can be categorized into permanent water, seasonal water, and historic floodwater. The DEM data used in this study are derived from the European Space Agency's Copernicus DEM, with a spatial resolution of 30 m (Li et al., 2022).

The Z-score method is a statistical anomaly detection method that has been proven effective for monitoring newly formed surface water bodies. Compared with other anomaly detection approaches, the Z-score method relies on commonly used statistical measures – mean and standard deviation (SD) – which offer strong interpretability. Thresholds can be defined based on the desired confidence level, providing solid theoretical support for better understanding the data and results (Fernandez et al., 2016; Cheadle et al., 2003). The Z-score method standardizes data using the mean and std, and it assumes that the data follow a normal distribution. If the data deviate significantly from normality, the reliability of the Z-score results may be compromised. Experiments conducted by DeVries et al. (2020) in various study areas demonstrated that backscatter characteristics of natural areas during the same season can generally be considered normally distributed, and their water body monitoring based on the Z-score method achieved an overall accuracy of 90 %. To verify whether the backscatter intensity of building areas in time-series SAR images follows a normal distribution, this section employs kurtosis and skewness as statistical indicators.

Kurtosis is a statistical measure used to describe the shape of a data distribution. It indicates the degree to which a dataset contains outliers and tail risk. A normal distribution has a kurtosis value of 3. Subtracting 3 from the kurtosis yields excess kurtosis, which equals 0 for a perfectly normal distribution. The formula for calculating excess kurtosis is (DeCarlo, 1997): 1Kurtosis=n(n+1)∑xi-μ4(n-1)(n-2)(n-3)σ4-3(n-1)2(n-2)(n-3)-3 where xi represents the elements in the array and n is the sample size, while μ and σ denote the mean and standard deviation of the array, respectively. The sign of the excess kurtosis indicates whether the shape of the distribution is sharper or flatter compared to a normal distribution. A positive excess kurtosis suggests a distribution with heavier tails and a higher likelihood of outliers, whereas a negative excess kurtosis indicates fewer outliers and a flatter distribution.

Skewness is a statistical measure that quantifies the degree of asymmetry in a data distribution. For example, a normal distribution has a skewness value of 0. In cases of asymmetric distributions, if the tail is longer on the left side, it is referred to as left-skewed (negative skewness); if the tail is longer on the right, it is right-skewed (positive skewness). Skewness is calculated by summing the cubed deviations from the mean and normalizing by the standard deviation. The formula for skewness is (Groeneveld and Meeden, 1984) : 2Skewness=1n×∑i=1Nxi-μ3σ3 where xi represents the elements in the array and n is the sample size, while μ and σ denote the mean and standard deviation of the array, respectively. Skewness can be either positive or negative: right-skewed data has positive skewness, while left-skewed data have negative skewness. Skewness can be used to assess whether a dataset follows a normal distribution or another specific distribution, thereby providing a basis for statistical analysis or modeling.

To quantitatively assess the normality assumption underlying the Z-score transformation, we calculated the median skewness and median excess kurtosis of the same-orbit Sentinel-1 backscatter time series for representative natural and urban land-cover classes during the non-flood reference period (Table 1). The selected classes included cropland, bare soil, open water, low-rise buildings, and high-density buildings. The diagnostic results indicate that the temporal backscatter distributions were not perfectly Gaussian. For the individual VV and VH backscatter channels, the median skewness ranged from 0.734 to 1.740, and the median excess kurtosis ranged from 0.048 to 2.596. All classes showed positive skewness, indicating that the non-flood backscatter time series generally had a right-skewed distribution. The deviation from normality was relatively small for urban VH backscatter, especially in low-rise buildings, where the median skewness and excess kurtosis were 0.734 and 0.048, respectively, and in high-density buildings, where they were 0.953 and 0.560, respectively. This suggests that VH backscatter in built-up areas was comparatively stable during the reference period. In contrast, cropland showed larger deviations, particularly in VH polarization, with a median skewness of 1.740 and a median excess kurtosis of 2.596. This is likely related to crop phenology, vegetation structure changes, and rainfall-induced surface moisture variations. The VV / VH ratio showed stronger departures from normality than the individual VV and VH channels, with median skewness ranging from 1.359 to 2.921 and median excess kurtosis ranging from 1.443 to 3.330. The largest skewness was observed for open water in VV / VH, indicating that ratio-based polarimetric variables can be more sensitive to low backscatter values and surface roughness variations. Therefore, the Z-score confidence interpretation in this study is applied primarily to the independently standardized VV and VH backscatter channels rather than to the VV / VH ratio. These results show that the normality assumption is only approximately satisfied. Accordingly, the Z-score thresholds should not be interpreted as exact pixel-level flood probabilities in all land-cover types. Instead, they are used as standardized anomaly levels that retain statistical interpretability under approximate normality. This interpretation is particularly important for vegetation-covered areas and mixed urban pixels, where skewness and excess kurtosis indicate stronger deviations from a Gaussian distribution.

Regarding the effect of sample size on skewness in normally distributed data, Doane and Seward (2011) provided skewness coefficient thresholds corresponding to a 90 % confidence level. Specifically, when the sample size is 40, the skewness should fall within the range of -0.594 to 0.594 (Doane and Seward, 2011). The Z-score values – accounting for differences in polarization mode, orbit parameters, and imaging mode – were then calculated from the mean and standard deviation of the time-series SAR images using the following formula: 3Zfloodpq,o,m=σfloodpq,o,m-meanσpq,o,mSDσpq,o,m where Z represents the Z-score value, pq denotes the polarization mode (VV or VH), o indicates the orbit number, and m refers to the imaging mode. The normalized backscatter coefficient is standardized using the mean and standard deviation of the corresponding same-orbit reference time series. Under approximate normality, the Z-score can be associated with a nominal confidence level indicating the statistical extremeness of the flood-period observation relative to the reference distribution. For example, when |Z| exceeds 1.96, the observation lies outside the central 95 % interval of the reference distribution. However, because Sentinel-1 backscatter may deviate from normality in phenologically dynamic vegetation and mixed urban pixels, the Z-score is interpreted in this study primarily as a standardized anomaly magnitude rather than an exact pixel-level flood probability. The expected anomaly direction differs by surface type: open-water flooding in natural areas usually produces negative anomalies, whereas urban inundation can produce positive anomalies through enhanced double-bounce scattering.

Using single-polarization SAR imagery alone yields limited accuracy in flood monitoring. Therefore, this study incorporates both VV and VH polarizations to provide multi-dimensional observational information and enhance the reliability of the flood mapping results. For example, during periods of heavy rainfall, rain and strong winds can increase water surface roughness, potentially leading to Bragg scattering dominance. This can strengthen the backscatter in co-polarized channels and obscure the presence of water bodies. In such cases, cross-polarization is less affected by surface roughness, offering additional evidence for flood delineation (Clement et al., 2018). In natural areas, the reference-baseline images used in this study were acquired 2–3 months prior to the flood event. If certain regions were continuously covered by water during this period, anomaly detection methods would struggle to identify them as flooded. To address this, auxiliary permanent surface water data from the GSW v1.4 dataset was incorporated to ensure the accuracy of flood extent delineation (Pekel et al., 2016). This section adopts the conceptual framework of decision tree classification, and by simultaneously considering dual-polarization information and surface water data, proposes a novel flood monitoring classification tree.

In this study, Eq. (3) was applied pixel by pixel to both VV and VH images to generate their respective Z-score maps. Because VV and VH polarizations have different physical sensitivities to surface roughness, vegetation structure, and urban double-bounce scattering, the use of identical Z-score thresholds for the two polarization channels requires further clarification. In this study, the same nominal threshold was applied to VV and VH only after each polarization channel had been standardized independently using its own same-orbit temporal mean and standard deviation. Therefore, a given threshold value represents the same standardized anomaly magnitude within each polarization, rather than the same absolute backscatter change in dB. This treatment reduces the direct influence of the different dynamic ranges of VV and VH and allows the two polarizations to be combined within a statistically consistent anomaly-detection framework. For natural areas, floodwater generally causes a negative backscatter anomaly due to specular reflection; therefore, pixels with the VV and VH Z-score values lower than the negative threshold were identified as flood candidates. For building areas, urban inundation can enhance double-bounce scattering between water surfaces and vertical structures; therefore, pixels with the VV and VH Z-score values higher than the positive threshold were identified as urban flood candidates.

The classification tree algorithm begins by determining the land cover type of each pixel. In natural areas, the first step is to identify whether a pixel represents a seasonal water body based on its water occurrence frequency. Typically, a pixel is classified as a seasonal water body if its occurrence value exceeds 25, meaning the pixel is identified as water in more than 25 % of the annual satellite observations. Such pixels are labeled in blue and excluded from further classification. For non-seasonal water bodies, if both the VV and VH Z-score values of a given pixel fall below a predefined threshold (th_n), the pixel is assigned a dual-polarization flood label. If only one of the polarizations falls below the threshold (th_n), the pixel is assigned a single-polarization flood label. In urban (building) areas, flood labels are similarly assigned based on whether the Z-score values of VV and VH are below a separate threshold (th_s). The same rule-based logic is applied to determine single- or dual-polarization flood classifications. It is important to note that in building areas, the color coding for flood labels differs from that used in natural areas.

The same numerical thresholds are used for the independently standardized VV and VH Z-score maps because each polarization is normalized by its own temporal mean and standard deviation. Thus, an identical Z-score threshold represents a comparable standardized anomaly level within each polarization. Nevertheless, VV and VH have different sensitivities to wind-roughened water, vegetation structure, and urban double-bounce scattering. The GSW occurrence threshold of 25 % was selected to remove recurrent seasonal/permanent water while retaining rarely inundated floodplain pixels. To evaluate the sensitivity of this choice, occurrence thresholds of 10 %, 25 %, 50 %, and 75 % were tested. Before classification, the 30 m GSW layer was co-registered to the Sentinel-1 processing grid. Because resolution mismatch and geolocation errors mainly affect narrow river margins, an additional uncertainty assessment excluded a one-pixel buffer around persistent-water boundaries during accuracy evaluation.

In SAR-based flood monitoring, even when filtering algorithms are applied during the preprocessing stage, speckle noise cannot be completely eliminated. As a result, the final flood mapping output often contains scattered misclassified pixels within large, homogeneous land cover areas. These fragmented misclassifications negatively impact the overall accuracy and may reduce the reliability of the flood monitoring results. To address this issue, this study adopts a method known as pixel aggregation. This technique specifically targets small misclassified regions. In particular, if a flood-classified area consists of four or fewer connected pixels, these pixels are reclassified as non-flood. The goal is to minimize the impact of noise-induced misclassification and improve the overall classification accuracy. This step serves as a post-processing refinement of the threshold-based classification results. Other commonly used post-processing methods include morphological operations such as erosion and dilation based on opening and closing. Through such post-processing techniques, small fragmented errors can be effectively removed, thereby enhancing the accuracy of flood monitoring outcomes (Borrelli et al., 2020). To avoid over-smoothing, this connected-pixel threshold was evaluated using 0, 2, 4, and 8 pixels. The ≤ 4-pixel rule was retained as the operational setting because it suppresses isolated speckle-induced false alarms while preserving the main flood pattern; however, it may remove very small genuine puddles or narrow drainage features.

In this study, Sentinel-2 optical images and high-spatial-resolution unmanned aerial vehicle (UAV) optical images were used as validation data sources for flood extraction results in natural areas and urban (building) areas, respectively. The time difference between the Sentinel-2 images and the corresponding Sentinel-1 acquisition was 8–9 h, while the UAV images had a 2 h time difference. For water body extraction, the modified normalized difference water index (MNDWI) utilizes the green and short-wave infrared (SWIR) bands, which enhances the contrast between building areas and water bodies. This reduces confusion between the two and improves classification accuracy. The expression is as follows (Xu, 2006): 4MNDWI=Green-SWIRGreen+SWIR where, in the Sentinel-2 images, the green and SWIR bands correspond to bands 3 and 11, respectively. Since MNDWI is a normalized index, it also performs well in detecting water bodies located in shadowed areas.

In this study, high-confidence and medium-confidence flood pixels are both classified as flood pixels during accuracy assessment. Therefore, binary classification metrics are used for evaluation. The confusion matrix for binary classification is a 2 × 2 matrix, with the four elements being: True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). Common evaluation metrics based on these four elements include overall accuracy (OA), user accuracy (UA), and producer Accuracy (PA). In flood monitoring, the flood class requires special attention, and prediction accuracy may not always be the primary concern – particularly in cases where class distribution is highly imbalanced. The critical success index (CSI), which considers both PA and UA, is a more suitable metric for flood monitoring as it effectively reflects both the accuracy and stability of predictions. Therefore, this study adopts four evaluation metrics, including CSI. Their formulas are as follows:

To quantify uncertainty caused by the 8–9 h time difference between Sentinel-1 and Sentinel-2 acquisitions, the two available Sentinel-2-derived water masks before and after the SAR acquisition were used to define a temporal uncertainty envelope. Pixels that changed between the two optical masks were treated as flood-transition pixels. Accuracy metrics were then recalculated after excluding these dynamic pixels, allowing the effect of temporal mismatch on CSI and OA to be estimated.

For urban validation, UAV imagery was used as a high-confidence reference because of its high spatial resolution and approximately 2 h temporal difference from the Sentinel-1 acquisition. However, the UAV images cover only part of the urban domain. Therefore, the UAV-based accuracy assessment should be interpreted as a local high-resolution validation rather than a complete city-wide assessment. 5PA=TPTP+FN6UA=TPTP+FP7OA=TP+TNTP+FP+FN+TN8CSI=TPTP+FP+FN

When using the Z-score method for flood monitoring, it is necessary to set a threshold to distinguish between anomalous (flooded) and non-anomalous (non-flooded) pixels. However, unlike traditional thresholding methods, thresholds based on Z-scores have a clear statistical confidence level. To compare the accuracy of flood monitoring results under different confidence levels, this section defines three Z-score thresholds: -2, -2.5, and -3. According to the standard normal distribution table, these thresholds correspond to confidence levels of 95.4 %, 98.8 %, and 99.7 %, respectively. The Z-score images generated using reference-baseline images and the time-series SAR flood monitoring algorithm were clipped to match the coverage area of the validation images. The Z-score values were then classified using three thresholds, as shown in Fig. 5. It is important to note that seasonal water bodies were excluded during the accuracy assessment. This is because such water bodies consistently exhibit low backscatter intensity in time-series SAR images, whereas time-series monitoring primarily aims to highlight newly emerged surface water.

In Fig. 5, different colours are used to represent the flood monitoring results under different Z-score thresholds. As shown, lowering the confidence level introduces more speckled flood pixels but also captures more water bodies. For instance, noticeable pink pixels appear in the northern and eastern parts of Weihui under less strict thresholds, whereas using a threshold of -3 results in more missed detections in these areas. A comparison between Fig. 5a and Fig. 5b reveals a significant difference between the flood extent identified by Sentinel-2 on the 26th and that detected by SAR on the 27th, particularly in the floodplains upstream and downstream of the Wei River. These differences are also reflected in the accuracy assessment, where the CSI and PA show discrepancies of up to 17 % and 23 %, respectively. The OA shows a slight increase, primarily because the flood and non-flood classes in Weihui are unevenly distributed, resulting in a high proportion of TN in the confusion matrix.

A comparison between the optical water extraction results from 26 and 31 July reveals significant changes in the flood extent within Weihui City during this period. The flood areas in the upstream and midstream regions contracted, while the flooded areas in the downstream region expanded. As shown clearly in Fig. 6, the monitoring results from 27 July are highly consistent with the ground truth from 31 July in the downstream floodplain, whereas there is a notable difference when compared to the July 26 ground truth. This highlights the dynamic nature of flood disasters. When evaluating the accuracy of flood monitoring algorithms, it is crucial to consider both timeliness – the validation images should be as close in time as possible to the SAR acquisition – and sample size, using as much validation data as possible to ensure accurate and reliable assessment of the algorithm's performance. In the accuracy evaluations using the two Sentinel-2 scenes as ground truth, the proposed SAR-based flood monitoring algorithm achieved over 92 % overall accuracy in both cases and showed strong consistency with the water extent derived from both the 26 and 31 July ground truths. Overall, while lowering the confidence level introduced more speckled flood pixels, which led to a noticeable decline in UA, the CSI – which places greater emphasis on the correct identification of flood areas – increased to varying degrees. Therefore, a Z-score threshold of -2 can be considered appropriate for flood detection in natural areas. At the 95 % confidence level, both study areas achieved approximately 90 % OA and CSI values above 60 %. In conclusion, the method proposed in this study demonstrated high accuracy within the study areas, indicating strong reliability and stability for flood monitoring applications.

In urban flood monitoring, it is also necessary to determine thresholds for identifying anomalies. This study found that the frequency distribution curves of Z-score images in building areas do not exhibit a bimodal pattern, indicating that the Otsu method (maximum inter-class variance) is not suitable for determining accurate thresholds in this context. Therefore, this section adopts the same confidence-level-based thresholds used for natural areas to monitor urban flooding, with threshold values set at 2, 2.5, and 3. In natural areas, using VV and VH polarizations from Sentinel-1 images helps mitigate interference from rough water surfaces, yielding better monitoring performance than single-polarization approaches. However, few studies have applied Sentinel-1 images for urban flood detection. Thus, this section first evaluates the performance of VV, VH, and dual-polarization combinations under the same threshold conditions in monitoring urban flooding.

The UAV images of Weihui City offer not only ultra-high spatial resolution and oblique viewing advantages, but also excellent timeliness. Given the 2-hour difference between the UAV and SAR acquisition times, it can reasonably be assumed that the flood extent remained largely unchanged during this interval. In this section, this validation dataset is used for a detailed analysis of the urban flood detection results based on time-series Sentinel-1 SAR images (Fig. 7). This study found that lower threshold values result in a significantly higher number of detected flood pixels compared to higher thresholds. In many areas, flood pixels identified at lower thresholds effectively filled the gaps and voids present in the higher-threshold results. Overall, the flood regions identified using lower thresholds exhibited better spatial continuity. Aside from a slight decline in UA, all other accuracy metrics showed improvements to varying degrees. Notably, the CSI – a key indicator of flood detection performance – achieved a 7.88 % increase in accuracy.

By comparing the identified flood regions with the reference data (Fig. 8), we observed substantial discrepancies in several building areas. This section focuses on two representative regions where the algorithm failed to detect urban inundation effectively even under a relatively low threshold, indicating that the expected flood-induced backscatter enhancement was weak. Orthorectified images alone are insufficient to explain this phenomenon. Therefore, the UAV oblique views shown in Fig. 8a and b were examined to better understand the local building and inundation conditions.

The UAV images show that these areas are low-lying and are dominated by single-story or low-rise residential buildings. In these locations, floodwater nearly reached the rooftops, substantially reducing the relative height difference between buildings and the surrounding water surface. Under such conditions, the radar return is dominated mainly by rooftop surface scattering, while wall–water double-bounce scattering becomes weak or absent. As a result, severe urban inundation may still be missed by intensity-based SAR detection methods.

This failure mechanism indicates that the building–water height difference is a key factor controlling the detectability of urban inundation in SAR imagery. When floodwater remains substantially lower than building height, wall–water double-bounce scattering can be enhanced, increasing backscatter intensity and facilitating flood detection. In contrast, when floodwater approaches the roof level of low-rise buildings, especially in dense residential areas with narrow building spacing, the side-looking SAR geometry may not produce a clear double-bounce response. This leads to weak flood-related backscatter anomalies and increases the risk of missed detection.

High-spatial-resolution DEM or digital surface model (DSM) data, building footprints, and building-height information would be useful for identifying such failure-prone areas by estimating local building–water height differences. However, the Copernicus DEM used in this study has a spatial resolution of 30 m and cannot resolve individual buildings, narrow streets, or local elevation differences. Therefore, the proportion of building areas affected by this failure mechanism was not quantitatively estimated in the present study. This limitation is further discussed as a future research direction in Sect. 5.

To further evaluate the reliability of the proposed method, we compared it with four baseline methods, including VV-only Z-score detection, VH-only Z-score detection, simple image differencing, and Otsu thresholding (Table 2). All methods were evaluated using the same Sentinel-1 acquisitions and validation masks to ensure a fair comparison.

In natural areas, Otsu thresholding achieved the highest CSI and OA, with values of 63 % and 92 %, respectively. This result indicates that histogram-based thresholding can perform well in open floodplain environments where flooded and non-flooded pixels show relatively clear backscatter separation. The proposed dual-polarization Z-score tree achieved slightly lower but comparable accuracy in natural areas, with a CSI of 60 % and an OA of 90 %. Compared with VV-only Z-score detection, the proposed method improved CSI and OA by 9 and 10 percentage points, respectively. Compared with VH-only Z-score detection, the improvements were 4 percentage points for both CSI and OA. Compared with simple image differencing, the proposed method improved CSI by 4 percentage points and OA by 5 percentage points. These results indicate that the proposed dual-polarization framework retains competitive performance in natural floodplain areas while reducing the limitations of single-polarization and simple change-detection methods.

The advantage of the proposed method is more evident in building areas. The proposed method achieved the highest CSI and OA, reaching 62 % and 73 %, respectively. In comparison, VV-only Z-score detection achieved a CSI of 52 % and an OA of 63 %, while VH-only Z-score detection achieved a CSI of 54 % and an OA of 65 %, respectively, indicating that single-polarization information is insufficient to fully capture the complex scattering response of urban inundation. Simple image differencing performed worse in building areas, with a CSI of 45 % and an OA of 56 %, because non-flood-related backscatter changes caused by rainfall, surface wetness, and urban structural heterogeneity can be incorrectly interpreted as flood-induced changes. Otsu thresholding showed the largest performance degradation, with CSI and OA decreasing to 39 % and 50 %, respectively. This confirms that Otsu thresholding is unstable in building areas where the Z-score distribution does not exhibit a clear bimodal pattern.

The Sentinel-1 Global Flood Monitoring product is an operational near-real-time flood mapping service of the Copernicus Emergency Management Service. It processes incoming Sentinel-1 SAR acquisitions globally and derives flood extent using an ensemble of three independent automated flood-mapping algorithms. The final observed flood extent is generated by a majority-decision ensemble and is accompanied by reference water, exclusion mask, likelihood, and advisory-flag layers. Although GFM provides global near-real-time flood information and a historical Sentinel-1 archive from 2015 onward, its effective availability and reliability depend on the Sentinel-1 acquisition schedule and on exclusion-mask conditions such as topographic distortion, radar shadow, dense vegetation, building areas, and low-sensitivity surfaces. Therefore, for urban flood mapping in Weihui City, the Sentinel-1 Global Flood Monitoring product appears to be ineffective. The accuracy of flood extraction from the natural areas of this product is slightly lower than that of the method used in this study.

Overall, the proposed method provided the most balanced performance across both natural and urban environments. Its mean CSI and OA across the two area types were 61 % and 81.5 %, respectively, which were higher than those of all baseline methods. More importantly, the proposed method had the highest worst-case performance, with a minimum CSI of 60 % and a minimum OA of 73 % across the two environments. In contrast, although Otsu thresholding performed best in natural areas, its CSI and OA decreased sharply in building areas. These results demonstrate that the proposed dual-polarization Z-score classification tree is not only accurate in open floodplain areas but also more reliable and transferable for complex urban inundation mapping.