Flood Susceptibility Mapping Using Machine Learning and Geospatial-Sentinel-1 SAR Integration for Enhanced Early Warning Systems

Highlights

What are the main findings?

• A multi-year flood inventory was generated from Sentinel-1 SAR imagery (2018–2023) using Google Earth Engine, capturing repeated ponding occurrences as inputs for the target of susceptibility modeling.
• Flood susceptibility maps were developed using both a statistical model (FR) and machine learning models (RF, XGBoost, CNN), with model performance assessed through AUC and feature interpretability evaluated with SHAP and validated with available high-risk locations monitored by early warning flood sensors.

What is the implication of the main finding?

• The integration of SAR-based flood inventory with geospatial factors provides a robust framework for identifying high-frequency flood-prone areas in Jefferson County, TX, serving as a representative example for data-scarce regions.
• The methodology supports data-driven flood risk management by offering accurate, interpretable, and transferable tools that can inform planning and adaptation strategies in other flood-prone regions.

Abstract

This study presents a comprehensive framework for flood susceptibility mapping by integrating geospatial factors with both statistical and machine learning models. Thirteen Flood-related factors, including DEM, slope, TWI, NDVI, etc., are extracted as features of models, and historical flood data derived from Sentinel-1 SAR from 2018 to 2023 are used as the target variables of the models. These datasets are analyzed using a frequency-based statistical model and three machine learning models, including Random Forest, XGBoost, and CNN, to generate flood susceptibility maps. The performance of each model is evaluated through AUC; and SHAP scores are separately generated for Machine learning (ML) models to explain each feature contribution in the ML model. The generated susceptibility maps are validated by high-flood-risk locations monitored by flood sensors, BLE inundation models, and flood-prone areas suggested by the Local Community Task Force. The results indicate that the XGBoost model outperforms all other models, with an AUC of 0.92 and demonstrates the highest alignment with recommended high-flood-risk locations, while the frequency-based statistical model showed the weakest performance with an AUC of 0.65. SHAP value graphs highlight the elevation, slope, and TWI as the most influential features across all models. The susceptibility maps generated by the machine learning model show strong agreement with the BLE map and high-flood-risk areas identified by the local Community Task Force.

1. Introduction

Floods are natural events that regularly occur in the United States, including hurricanes and tornadoes, on small and regional scales, which the country experiences every year [1,2,3,4,5]. According to the Federal Emergency Management Agency (FEMA), flooding ranks as the foremost natural hazard in the United States [6]. Furthermore, the National Weather Service (NWS) reports that after heat waves, floods are the second leading cause of weather-related fatalities in the United States [7].

Over the last 50 years, thousands of flood events have been recorded, since 1975, and an average of 125 deaths annually, totaling roughly 5000–6000 fatalities nationwide, plus several thousand injuries (e.g., 146 deaths and 60 injuries in 2021 alone). Texas is the nation’s most flood-prone state, enduring hundreds of floods and over 70 major federal flood disasters since 1975; it also leads the country in fatalities, with an estimated 800+ deaths over the past 50 years and likely hundreds of injuries [8,9]. Southeast Texas, especially Jefferson County, is among the hardest-hit areas due to tropical storms and Gulf rainfall, suffering dozens of flood-related deaths over recent decades. Catastrophic events such as Hurricane Harvey (2017), which dumped a U.S. record 60.58 inches of rain in Jefferson County and caused 68 deaths statewide, and Tropical Storm Imelda (2019), which killed 5 people (3 in Jefferson County) and forced hundreds of rescues, highlight the extreme vulnerability of this region to deadly and destructive flooding [10,11].

The National Hurricane Center (NHC) officially recognized Hurricane Harvey as the most significant tropical cyclone rainfall event in U.S. history and the second most destructive hurricane, after Hurricane Katrina, in terms of losses [12,13]. This powerful storm impacted Southeast Texas from late August to September 2017, resulting in the recorded accumulation of over 60 inches of rainfall in the vicinity of Nederland and Groves, Texas [14]. Two years after Hurricane Harvey, Tropical Storm Imelda happened in Southeast Texas in September 2019, marking the fourth time in a decade that flooding closed Interstate I-10. In order to improve cooperation in handling the region’s regular floods, the Southeast Texas Flood Coordination Study (SETxFCS) was established in November 2019. The Flood Coordination Study project involves managing over 80 low-cost flood sensors across seven counties in coordination with DHS S&T. Tasks include installation support, asset management, elevation mapping, and establishing alert thresholds to enhance data visualization on OneRain platforms [15,16].

Flood Susceptibility Mapping is crucial for mitigating flood hazards by identifying vulnerable areas using Remote Sensing (RS) and Geographic Information System (GIS) tools. The increasing frequency and intensity of floods, driven by climate change and land-use changes, emphasize the need for effective flood susceptibility assessments. In this paper, the term flood susceptibility refers to all areas where water ponding has occurred. While ponding susceptibility may be a more precise term for this context, since the target values used in all modeling are water pixel data derived from Sentinel-1 SAR imagery, the broader and more widely accepted term flood susceptibility is used for clarity and general applicability. Flood susceptibility is determined by various physical factors such as digital elevation model (DEM), slope, soil type, land use/land cover (LULC), topographic wetness index (TWI), etc. The selection of these conditioning factors depends on the spatial scale of the study. Large-scale analyses use fewer factors for consistency, whereas local-scale studies incorporate more location-specific data for accuracy. Geographic Information Systems (GIS) play a key role in integrating these factors to create flood susceptibility maps [17,18,19,20,21,22].

Flood Susceptibility Mapping methodologies can be categorized into three main groups:

Hydrological Methods: Hydrological methods for flood susceptibility mapping include models like SWAT, WetSpa, and HYDROTEL, which simulate flood events by analyzing water flow dynamics and considering factors such as land use, soil type, topography, and climate data. These approaches are powerful as they capture complex rainfall–runoff interactions, integrate with climate change models, and can be combined with decision-making tools like AHP or TOPSIS to generate robust flood zoning maps. However, they also have weaknesses, including high data demands, reduced accuracy at small or local scales, and uncertainties in downscaling climate models such as GCMs, which limit their reliability in data-scarce regions [23,24].

Statistical and Knowledge-Based Methods: Approaches such as Weights of Evidence (WoE), Logistic Regression (LR), Frequency Ratio (FR), and Analytic Hierarchy Process (AHP) are commonly used in GIS to assess flood susceptibility based on conditioning factors. Multi-Criteria Decision Analysis (MCDA), particularly AHP, integrates socio-economic and environmental indicators for decision-making but faces limitations in handling uncertainties and nonlinear dynamics. These complexities are better addressed using physical models or advanced machine learning techniques, though they require extensive data, computational resources, and expert knowledge [25,26].

Machine Learning Methods: Machine learning (ML) methods have emerged as a reliable and objective approach for flood susceptibility and risk assessment, especially under the growing pressures of rapid urbanization and climate change. Techniques such as Artificial Neural Networks (ANNs), Decision Trees (DTs), Random Forests, XGBoost, Convolutional Neural Networks (CNNs), and Support Vector Machines (SVMs) can capture complex nonlinear flood patterns without relying on traditional statistical assumptions. Compared to historical statistics, scenario simulation, and multi-criteria decision analysis, ML models provide more flexible and data-driven predictions, though they often require large datasets, careful parameter tuning, and considerable computational resources [27,28,29,30,31,32,33,34,35].

Statistical models rely on past disaster records to estimate risk and often provide results consistent with observed events [36]. However, they require long-term historical data, have limited flexibility, and cannot easily adapt to rapidly changing urban and climatic conditions [37]. In contrast, machine learning models automatically learn flood risk patterns from diverse datasets and can capture nonlinear relationships that statistical methods often miss [38]. ML models are less dependent on large historical records, require less preprocessing than hydraulic models, and provide more objective and transferable assessments [39]. Numerous studies demonstrate that ML models outperform traditional statistical approaches in prediction accuracy and scalability, making them a superior choice for modern flood risk assessment [40].

This study presents different machine learning techniques to generate a flood susceptibility map and also compares it with a frequency-based statistical model. The FR method is particularly effective in GIS-based environments, utilizing remote sensing (RS) data for accurate flood hazard assessment and mapping.

The study focuses on Jefferson County, Texas, which faces severe flash floods during the year. Jefferson County is a critical area due to its combination of urban, suburban, and rural landscapes, which makes it particularly vulnerable to localized flooding and high-frequency ponding.

Table 1. The table of acronyms.

Abbreviation	Definition	Abbreviation	Definition
AHP	Analytic Hierarchy Process	MCDA	Multi-Criteria Decision Analysis
ANN	Artificial Neural Network	NDVI	Normalized Difference Vegetation Index
AUC	Area Under the Curve	NHC	National Hurricane Center
BSA	Bivariate Statistical Analysis	NLCD	National Land Cover Database
BLE	Base Level Engineering	NRCS	Natural Resources Conservation Service
BM	Benchmark (sensor ID, e.g., BM27)	NWS	National Weather Service
CNN	Convolutional Neural Network	PRISM	Parameter-elevation Regressions on Independent Slopes Model
DEM	Digital Elevation Model	RF	Random Forest
DHS	Department of Homeland Security	ROC	Receiver Operating Characteristic
DOE	Department of Energy	RS	Remote Sensing
DT	Decision Tree	SAR	Synthetic Aperture Radar
ECDF	Empirical Cumulative Distribution Function	SETxFCS	Southeast Texas Flood Coordination Study
FEMA	Federal Emergency Management Agency	SHAP	Shapley Additive Explanations
FN	False Negative	SSURGO	Soil Survey Geographic Database
FP	False Positive	SVM	Support Vector Machine
FR	Frequency Ratio	SWAT	Soil and Water Assessment Tool
FPR	False Positive Rate	TNRIS	Texas Natural Resources Information System
GEE	Google Earth Engine	TN	True Negative
GIS	Geographic Information System	TP	True Positive
LULC	Land Use/Land Cover	TPR	True Positive Rate
LR	Logistic Regression	TWI	Topographic Wetness Index
ML	Machine Learning	WoE	Weights of Evidence
MLT	Machine Learning Technique	XGBoost	Extreme Gradient Boosting

presents the acronyms of the research. The objectives of this study are

• Identifying which area in Jefferson County has high-frequency ponding hot spots using Sentinel-1 Radar Imagery.
• Using the Sentinel-1 Radar Imagery as a target for Machine learning and statistical models to create a high-frequency ponding susceptibility.

The generated susceptibility maps will be validated by the high-flood-risk locations monitored by the sensors, the BLE risk map, and the high-flood-risk areas identified by the local community task force.

The study is novel in the method of developing a multi-year pluvial flood inventory map from Sentinel-1 SAR imagery (2018–2023) using Google Earth Engine. This method, based on repeated ponding occurrences, provides a flood inventory dataset for susceptibility modeling. By integrating this dataset with 13 geospatial conditioning factors, both a statistical model (FR) and machine learning models (RF, XGBoost, CNN) were developed. This study is also unique in that it not only validates the accuracy of the machine learning and statistical model itself through AUC but also validates its performance against high-risk areas monitored by flood sensors. Applied for the first time to Jefferson County, Southeast Texas, a region highly prone to flooding, the study provides practical, robust data for flood risk management, and the output maps identify high flood-susceptible areas and inform community-based decision-making.

2. Materials and Methods

2.1. Description of the Study Area

Jefferson County is in Southeast Texas at approximately 29.8165°N, 94.1514°W, part of the Beaumont-Port Arthur Metropolitan Statistical Area and the most populous of its four counties. According to the 2010 census, the county had 252,273 residents, while in 2020, this population had reached 256,526. Jefferson County includes nine cities: Beaumont (The largest city is Beaumont, which had a population of about 115,282 in 2020), Bevil Oaks, China, Groves, Nederland, Nome, Port Arthur (The second-largest city is Port Arthur, with 56,039 residents), Port Neches, and Taylor Landing. According to the U.S. Census Bureau, the county has a total area of 1113 square miles (2878 km²), wherein 876 square miles (2800 km²) of it is land and 236 square miles (21%) of it is covered by water. The northern boundary of the county consists of Pine Island Bayou, while in the northeast lies the Neches River. In the east are Sabine Lake and the Sabine River. The southern part is all marshland down to the Gulf and contains Sea Rim State Park. The surrounding area constitutes a part of the Golden Triangle region of Southeast Texas, characterized by its industrial base, along with great cultural diversity. Jefferson County represents the Gulf Coastal Plain, which is characterized by flat to gently rolling terrain, wetlands, marshes, and bayous that present its ecological diversity. Elevations vary from sea level along the coast to a little higher inland. The climate is humid subtropical, classified as Cfa, with hot summers and mild winters. The average annual rainfall is about 60 inches (1524 mm), well distributed throughout the year. Temperatures range from an average low of 45 °F (7 °C) in January to a high of 92 °F (33 °C) in July and August (Figure 1).

2.2. Flood Susceptibility Factors and Data Preparation

The selection of flood conditioning factors is important in effective flood susceptibility mapping, but no universally applicable scheme is available. The most relevant and most utilized factors based on the recent studies have been used in this work. A total of 13 geospatial maps were delineated to generate a flood susceptibility map, including the Elevation Model (DEM), slope, Topographic Wetness Index (TWI), Normalized Difference Vegetation Index (NDVI), Soil type, Rock unite, land-use/land-cover (LULC), Depression areas, Soil Hydrologic group, Average Precipitation, Distance from major streams, Distance from minor streams, and distance from roads (Figure 2 and Figure 3 and

Table 2. The table of acronyms of the study.

Factor	Data Types	Resolution	Source
DEM	Raster	30 m × 30 m	USGS Elevation DATA
Slope	Raster	30 m × 30 m	DEM
TWI	Raster	30 m × 30 m	DEM
NDVI	Raster	30 m × 30 m	Landsat 8 imagery
Soil type	Polygon		Soil Survey Geographic Database 2.3.2 [48]
Rock Unite	Polygon		RockUnitPoly250K, Texas (TNRIS) Geologic Data
LULC	Raster	30 m × 30 m	NLCD 2021 Land Cover (CONUS)
Depression Areas	Raster	30 m × 30 m	DEM
Soil Hydrology group	Polygon		SSURGO Database
Average Precipitation	Polygon		PRISM (2018–2023)
Distance for streams and waterbodies	Polygon		Jefferson County Drainage District No. 6
Distance from Road	Polygon		TxDOT Roadways dataset

) [41,42,43,44,45,46,47]. A Sentinel-1 SAR map was generated with the Google Earth Engine code to retrieve data for pluvial flooded areas from 2018 to 2023 (Figure 4 and Figure 5).

2.2.1. Digital Elevation Model

DEM indicates elevation, where low-lying areas are more prone to water accumulation and flooding, while higher terrains facilitate runoff and reduce inundation risk [43]. High-resolution flood conditioning factors significantly enhance flood susceptibility mapping accuracy. Higher-resolution DEMs show more details and, consequently, they require more computational resources. Moderate-resolution DEMs (e.g., 30 m) offer a balance between accuracy and efficiency, making them widely used in flood susceptibility studies. The choice of DEM resolution also influences flood inventory mapping, which affects how well flood and non-flood areas are distinguished [49,50].

A 30 m resolution DEM is used in this study because of the time saving of the analysis, and likewise, previous research has shown that it has acceptable accuracy in generating flood susceptibility maps. in this study, 30 m DEM data were retrieved from USGS Elevation DATA.

2.2.2. Slope

Slope was derived from a Digital Elevation Model (DEM) using the Slope tool in ArcGIS Pro and calculated in percentage as a feature for both statistical and machine learning models. Areas with low slope values are more prone to flooding due to reduced runoff and poor drainage. Conversely, steeper slopes promote faster runoff, reducing water accumulation. Incorporating slope as a continuous geospatial layer allows for improved identification of flood-prone areas [51,52,53].

2.2.3. Topographic Wetness Index (TWI)

TWI reflects soil moisture potential, in which higher values present zones of water accumulation that are more likely to experience flooding [54]. TWI is calculated using the log-transformed ratio of the upslope area to the contour length (Equation (1)). In Equation (2), 0.5 is added to the denominator of the fraction to prevent division errors on flat terrain, which is used in this study, as Jefferson County is located on a flat plain area [43].

$$TWI=\mathrm{l}\mathrm{n}({\displaystyle \frac{\alpha }{tan\beta }})$$

(1)

$$TWI=\mathrm{l}\mathrm{n}({\displaystyle \frac{\alpha }{tan\beta +0.5 }})$$

(2)

where $\alpha$ is the cumulative upstream discharge at one point and β is the local slope angle.

2.2.4. Normalized Difference Vegetation Index (NDVI)

Areas with high NDVI (dense vegetation) promote infiltration and reduce runoff, while low NDVI (bare or urban surfaces) increases surface flow and flood potential [55]. The Google Earth Engine script was developed to process Landsat 8 imagery to generate the NDVI map of Jefferson County, TX. Images from 1 January 2024, to 31 December 2024, were filtered; only those with less than 1% cloud cover were selected and averaged. The NDVI was calculated using the near-infrared (B5) and red (B4) bands. In the end, the NDVI image was saved with a 30-m resolution [56,57,58].

2.2.5. Soil Type

Soil type plays a key role in flood susceptibility by controlling infiltration and runoff. Sandy soils with high permeability reduce flood risk by absorbing water, while clay soils with low permeability increase surface runoff and thus higher flood risk [43,59]. The soil type retrieved from the Soil Survey Geographic Database 2.3.2 [60], updated by 30 August 2024. Based on the SSURGO database, Jefferson County is mainly covered by clay and clay loam soils, which have low permeability and poor drainage. This increases surface runoff and makes the area more vulnerable to flooding during heavy rainfall.

2.2.6. Rock Unit

Geological formations control permeability, with permeable rocks like sandstone enhancing infiltration, whereas impermeable units such as shale or granite increase surface runoff and flooding likelihood [61,62]. In Jefferson County, the majority of areas are under the Qal and Qb rockunit classifications, which are 24.60% and 70.62% of the whole area, respectively. Based on geological mapping, the Qal is classified as a very high flood hazard area that is vulnerable to riverine floods, flash floods, and debris flow. Qb refers to flood-basin deposits from the Recent age, and Qal represents undifferentiated younger and older alluvium from the Quaternary age [63,64]. Rock Unit data driven from RockUnitPoly250K, Texas (TNRIS) Geologic Data [65].

2.2.7. Land Use Land Cover

Land Use Land Cover (LULC) significantly influences flood susceptibility by influencing surface runoff and infiltration [66,67]. While urban areas with impervious surfaces and uncovered land increase runoff and flood risk, areas with high vegetation density have a lower risk of flooding due to their ability to absorb water [68]. In this study, the Land Use Land Cover (LULC) from the NLCD 2021 Land Cover (CONUS) was downloaded and classified into agricultural land, barren land, developed areas, forested upland, Herbaceous, shrubland, wetlands, and water bodies, which each affect flood dynamics differently [69]. These classes help assess flood-prone areas by determining how land cover influences inundation. Urban expansion or deforestation can change LULC over time, which can alter flood patterns and the necessity of continuous monitoring. Studies have shown a strong correlation between LULC changes and flood occurrences, emphasizing the need for sustainable land-use planning to mitigate flood risks [70,71].

2.2.8. Depression Areas

A depression is a low-lying area surrounded by higher ground, where water accumulates during rainfall and is highly suffers from nuisance flooding [72]. Pluvial flooding in urbanized areas only occurs in depressions, which hold stormwater that exceeds the capacity of urban drainage systems. These are important in urban flooding because they can store rainwater, leading to localized flooding if drainage is insufficient. Depression is simply calculated by subtracting the filled DEM from the original DEM [73,74]. In this study, to enhance the sensitivity of the susceptibility model, all filled areas are considered depression areas [75].

Depression areas = Filled DEM − Original DEM

(3)

2.2.9. Soil Hydrology Group

The soil hydrology group is downloaded from SSURGO [48] database. Based on their potential for runoff, soils have been divided into four major groups by the Natural Resources Conservation Service (NRCS). Jefferson County is characterized by all four soil hydrologic groups A, B, C, and D in which about 91.68% of the land area is in soil hydrologic group D. Soil hydrologic group D indicates that soils having very slow infiltration rates when thoroughly wetted, consisting of mostly clayey soils with high swelling capacity or potential, soils with a high permanent water table, soils with claypan or clay layer at or near the surface, and shallow soils over nearly impervious materials. These soils have a very slow rate of water transmission. The rest of the groups in Jefferson County are B/D at 3.53%, C/D at 2.89%, and group A, which is 1.1%. Group B/D shows soils that naturally have a very slow infiltration rate due to a high water table but will have a moderate rate of infiltration and runoff if drained. Soil C/D shows that soils naturally have a very slow infiltration rate due to a high water table but will have a slow rate of infiltration if drained, and group A stands for soils consisting of deep, well-drained sands or gravelly sands with high infiltration and low runoff rates.

2.2.10. Average Precipitation

Areas with high rainfall receive more water input, which increases runoff and flooding, especially during extreme events like hurricanes or intense storms [76]. Based on the rain data downloaded from the Parameter-elevation Regressions on Independent Slopes Model (PRISM) website from 2018 to 2023, the study area experienced an average annual precipitation of 1695 mm. The northeast of Jefferson County, which also consists of Beaumont city, experienced rain higher than 1695 mm, a total of 19.5% of the entire area (Figure 2 and Figure 3j) [77].

Figure 2. Yearly average flood retrieved from PRISM.

Figure 2. Yearly average flood retrieved from PRISM.

Figure 3. Geospatial layers for delineating a flood susceptibility map. (a) Dem, (b) Slope, (c) TWI, (d) NDVI, (e) Soil Type, (f) Rock Unit, (g) LULC, (h) Depression, (i) Soil Hydrology, (j) Precipitation, (k) Distance from Major Streams, (l) Distance from Minor Streams, (m) Distance from Roads.

Figure 3. Geospatial layers for delineating a flood susceptibility map. (a) Dem, (b) Slope, (c) TWI, (d) NDVI, (e) Soil Type, (f) Rock Unit, (g) LULC, (h) Depression, (i) Soil Hydrology, (j) Precipitation, (k) Distance from Major Streams, (l) Distance from Minor Streams, (m) Distance from Roads.

2.2.11. Distance for Streams and Waterbodies

The possibility of flooding is enhanced when an area is near rivers and canals and decreases if it is farther away. For this reason, areas near rivers and canals are given a higher vulnerability weight in comparison with those farther away [52,78]. In this article, based on the previous findings of [41,79] streams are categorized into two groups, major and minor, based on their surface width, as their behaviors are different during floods, which results from differences in their floodplains. Small ditches, canals, and rivers with surface widths less than 10 m are considered minor streams, and the rest of the water bodies are considered major. For each major and minor stream, five classification categories are considered using the Euclidean Distance tool in ArcGIS Pro 3.0.0.

2.2.12. Distance from Road

The distance from roads is an independent variable that plays a crucial role in flood risk assessment. Areas closer to roads typically have more impervious surfaces, which accelerate runoff during rainfall events, increasing the likelihood of flooding. while areas farther from roads tend to have more permeable surfaces, allowing for greater water absorption and retention, thereby reducing flood risk [46,80]. The Euclidean Distance tool in ArcGIS Pro was employed for calculation, and the result was classified into five groups based on quantile (Figure 3m).

2.2.13. Sentinel 1 SAR Image

For flood extent mapping, spectral data from optical sensors like Landsat and backscatter data from synthetic aperture radar (SAR), such as Sentinel-1, are commonly used. Optical data are preferred in cloud-free conditions due to their strong correlation with open water [81,82,83]. A JavaScript code is developed for use in Google Earth Engine (GEE) to automatically retrieve water pixels from 2018 to 2019, using Sentinel-1 SAR imagery with the following Algorithm 1:

Algorithm 1. Flood Inventory Mapping Using Sentinel-1 SAR Imagery.

Input: Sentinel-1 SAR imagery (2018–2019), PRISM rainfall data Output: Water pixels map

(1) Obtain rainy and dry dates from PRISM to distinguish wet and dry days. (2) For each rainy day:   Select the nearest subsequent Sentinel-1 SAR image. (3) For each dry date: Select the nearest Sentinel-1 image to the corresponding wet image, ensuring ≥15 consecutive dry days before selection. (4) Apply a water mask with VV polarization threshold (<−13.05 dB) and perform speckle noise filtering. (5) Define water layers:   • waterWet: water pixels during wet day   • waterDry: water pixels during dry day   • waterDry.eq(0): pixels not water in dry day   • waterWet. and (waterDry.eq(0)): pixels wet in wet day AND not wet in dry day (6) Integrate all generated images for each year. (7) Retain only water pixels that appear repeatedly for ≥3 years to reduce noise and increase reliability.

The −13.05 dB VV threshold for Sentinel-1 was adopted from Islam and Meng [84], who calibrated it over Houston City, Texas. This threshold is considered for Jefferson County because both areas lie within the Gulf Coastal Plain of Southeast Texas, sharing low-lying topography, humid subtropical climate, and frequent pluvial (ponding) flooding conditions. In terms of proximity, Houston City (≈29.76°N, 95.37°W) is ~80 miles from the Jefferson County study area (≈29.82°N, 94.15°W).

Figure 4. VV Sentinel-1 image 2018–2023. (BM27 Location: 30.0423000, −94.1092000).

Figure 4. VV Sentinel-1 image 2018–2023. (BM27 Location: 30.0423000, −94.1092000).

A susceptible location is defined as a spatial cell that experiences multiple water pixel occurrences over the study period. Classifying cells based on a single water pixel observation can lead to overestimation, capturing many areas that may have been affected by isolated or incidental events. On the other hand, requiring flood occurrence in all six years is too stringent and may exclude areas that are moderately but consistently at risk. Therefore, those areas that experience at least three times water pixel occurrences during the study period are considered for the target of the model (Figure 4 and Figure 5–

Table 3. Number of flood cells based on flood occurrence.

Flood Occurrence (Times)	1	2	3	4	5	6
Number of water pixels	496,973	144,806	35,630	7342	1636	238

).

Figure 5. VV Sentinel-1 image 2018–2023—(a) All detected water pixels, (b) Water pixels that repeated at least 3 times.

Figure 5. VV Sentinel-1 image 2018–2023—(a) All detected water pixels, (b) Water pixels that repeated at least 3 times.

2.3. Methodology

The methodology of this study is divided into four main stages. Steps 1 and 2 fall under data preparation and are explained in detail in the flood susceptibility factors and data preparation section, while Steps 3 and 4 focus on model development and evaluation for flood susceptibility mapping (Figure 6).

Step 1. Data Preparation and Map Creation:

Thirteen geospatial conditioning factors were collected and processed to support flood susceptibility analysis. These included topographic (DEM, slope, TWI), environmental (NDVI, LULC, soil type, rock unit), hydrological (precipitation, soil hydrologic group, depression areas), and proximity factors (distance to roads, major streams, and minor streams). All datasets were standardized and converted into geospatial maps in a GIS environment. ArcGIS Pro was utilized for spatial delineation, ensuring accurate integration and processing of these layers. The feature space for modeling was ultimately constructed using these 13 geospatial layers (Figure 3).

Step 2. Flood Inventory Using Sentinel-1 SAR Imagery:

A flood inventory map was produced by processing Sentinel-1 SAR images from 2018 to 2023 in Google Earth Engine. Water pixels were extracted, and repeated occurrences (≥3 times) were classified as flood-prone to reduce incidental errors. The target variable was established as a binary classification label, where flooded areas were assigned a value of 1 and non-flooded areas a value of 0.

Step 3. Flood Susceptibility Mapping:

Flood susceptibility maps were generated through two complementary approaches:

• Statistical Model (Frequency Ratio, FR): The FR method was applied to calculate the relative likelihood of flooding for each class of conditioning factors, providing a baseline statistical assessment of susceptibility.
• Machine Learning Models: three machine learning models, Random Forest, XGBoost, and Convolutional Neural Networks (CNN), were employed to generate flood susceptibility maps. Random Forest and XGBoost were selected as tree-based algorithms, which are widely recognized as effective models for tabular datasets. Random Forest used as a robust ensemble method with relatively simple structure and interpretability, whereas XGBoost, as an advanced boosting algorithm, was included to capture complex nonlinear relationships and interactions feature space. The CNN model was also introduced as a deep learning architecture capable of learning local feature dependencies and nonlinear patterns across the geospatial predictors.

Step 4. Model Evaluation and Comparison:

The resulting susceptibility maps from both the statistical and machine learning approaches were compared against available high-risk locations, the Base Level Engineering (BLE) map, and input from the Local Community Task Force to assess their accuracy and practical relevance. Model performance was quantitatively evaluated using the Area Under the ROC Curve (AUC), while SHAP values were computed for the machine learning models to explain the contribution of each conditioning factor. This multi-layered validation ensured both predictive accuracy and interpretability of the models for flood risk management.

2.3.1. Model Development

Statistical Model

Frequency ratio calculations: The Frequency Ratio (FR) applies bivariate statistical analysis to link flood frequency with independent variables, reflecting the principle that past flood events can occur again under similar conditions. FR scores are quantitative, where values above 1 indicate a higher likelihood of flooding, while scores below 1 suggest lower susceptibility [85,86,87,88,89,90,91]. The frequency ratio (FR) is determined by analyzing the percentage of ponding occurrences (water-pixels) within each class of a given geospatial layer. Each layer is categorized into distinct classes by quantile or its own classification scheme, and for major and minor streams and distance from roads, classifications from previous studies were adopted, as already explained in the data preparation section. All maps are illustrated in Figure 3. First, the percentage of each class relative to the entire layer is computed. Next, the percentage of ponding events occurring within each class is calculated. The FR is then obtained by dividing the event percentage by the class area percentage and multiplying the result by 100, thereby assigning a weight to each class area.

Machine Learning

i. 

Model Preprocessing

Data Normalization: In this study, normalization was applied as a preprocessing step for the Convolutional Neural Network (CNN) model. Neural networks are sensitive to the scale of input variables, and normalization improves training stability and convergence by reducing the risk of exploding or vanishing gradients. For this purpose, the features used in the CNN were standardized using the Standard Scaler transformation. In contrast, Random Forest and XGBoost models were trained directly on raw feature values, as tree-based algorithms rely on threshold-based splitting rules and are inherently insensitive to feature scaling.

Training–Testing Split and Cross-Validation: The binary flood occurrence (water pixels) dataset used for machine learning was inherently imbalanced, with only 1.7% water pixels (positive class) compared to 98.3% non-water pixels (negative class). To ensure reliable evaluation of model performance, the dataset was divided into training and testing subsets using stratified sampling. This approach maintained the original proportion of flooded and non-flooded pixels in both sets, which was particularly important given the class imbalance. In addition, stratified k-fold cross-validation was applied during model training, ensuring that class distributions were preserved across folds and reducing the risk of biased evaluation.

Hyperparameter Tuning: The optimization procedure was carried out through three main steps. First, the dataset was divided into multiple stratified folds to preserve class proportions across both training and validation subsets. Next, an optimization function was defined to calculate the AUC score of each model configuration on the validation data. Finally, the optimum parameters were estimated by identifying the configuration that achieved the maximum AUC score among the randomly sampled hyperparameter sets.

ii. 

Models

Random Forest: Random Forest (RF) is an ensemble learning algorithm that constructs multiple decision trees using a bagging approach, making it effective for both classification and regression tasks. It leverages the advantages of decision trees, such as adaptability to missing values and robustness, while addressing their limitations, including sensitivity to noise and overfitting [92]. RF operates by training individual decision trees on randomly selected subsets of data and features, then aggregating their outputs through majority voting for classification or averaging for regression. In the RF algorithm, a random vector i_k (representing a conditioning factor) is independently generated for each tree and distributed across all trees. This results in a collection of tree-structured classifiers h(x,i_k) for input vector x. The trees are grown using both the training dataset and the random vector x, ensuring diversity among the classifiers. RF effectively handles feature correlations without requiring special adjustments and adapts well to various data distributions [93,94].

One of RF’s notable strengths is its interpretability; it can assess the importance of each feature by analyzing its contribution across decision trees. Additionally, RF’s hyperparameter tuning process is relatively simple, making it practical for real-world applications.

XGBoost: XGBoost (eXtreme Gradient Boosting) is a state-of-the-art boosting algorithm that improves upon traditional gradient boosting methods by optimizing computational efficiency, regularization techniques, and handling of missing values. Introduced by Chen and Guestrin [95], XGBoost has been widely applied in data science competitions and real-world applications due to its robustness and scalability. In this study, XGBoost is used to find the relationship between hydrological and geospatial features and flood susceptibility. XGBoost follows an additive model where a weak learner (a decision tree) is added at each iteration to minimize the loss function. The prediction at the K times iteration is given by

$${\hat{y}}_i^{\left(K\right)}={\hat{y}}_i^{(K-1)}+f_K\left(x_i\right)$$

(4)

where

• ${\hat{y}}_i^{\left(K\right)}$ is the predicted output after $K$ boosting rounds.
• $f_K\left(x_i\right)$ is the newly added decision tree.

The objective function consists of two main components: the training loss and the regularization term:

$$L=\sum _{i=1}^{n}l\left(y_i,{\hat{y}}_i\right)+\sum _{k=1}^K\Omega \left(f_k\right)$$

(5)

where

• $l\left(y_i,{\hat{y}}_i\right)$ represents the loss function (e.g., log-loss for binary classification).
• $\Omega \left(f_k\right)$ is the regularization term, which helps prevent overfitting and is defined as

$$\Omega \left(f_k\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda \Vert \omega {\Vert }^2$$

(6)

where

• $T$ is the number of leaves in the tree.
• $\gamma$ and $\lambda$ are regularization hyperparameters.
• $\omega$ represents leaf weights.

CNN: A Convolutional Neural Network (CNN) was employed as one of the deep learning architectures, as it can capture spatial patterns and feature interactions from multi-layer input data. For flood susceptibility mapping, environmental and topographic variables often interact in complex ways that are not easily represented by conventional methods, and CNNs provide a framework to model these spatial dependencies.

The CNN developed in this study was designed to take 13 geospatial predictor layers as input and process them through convolutional, pooling, and dense layers. The convolutional layers were used to detect patterns and interactions across the input features, while pooling layers reduced the size of the data and kept the most important information. The dense layers then combined these learned features and produced the final classification. The model was trained using the Adam optimizer with binary cross-entropy loss, and a sigmoid activation function was applied in the output layer for binary classification. The overall architecture of the CNN is presented in Figure 7.

2.3.2. Model Evaluation

• ROC Curve (AUC) metric

The Area Under the Curve (AUC) is derived from the Receiver Operating Characteristic (ROC) curve, which illustrates a model’s classification performance across different probability thresholds. The ROC curve plots the True Positive Rate (TPR) against the False Positive Rate (FPR), where TPR represents the proportion of actual positives correctly identified, and FPR represents the proportion of actual negatives incorrectly classified as positive. By adjusting the probability threshold used to convert predicted probabilities into binary outcomes, different true positive and false positive rates are produced, forming the ROC curve. The AUC score summarizes this curve into a single value between 0 and 1, where a value closer to 1 indicates excellent discriminative ability. In the context of flood susceptibility mapping, a higher AUC means the model is more capable of distinguishing between flood-prone and non-flood-prone areas over all threshold levels. The formulas for TPR and FPR are given by

$$TPR={\displaystyle \frac{TP}{TP+FN}}$$

(7)

$$FPR={\displaystyle \frac{FP}{FP+TN}}$$

(8)

where

▪

TP (True Positive) = number of ponding-prone areas correctly predicted as flood-prone

▪

FN (False Negative) = number of ponding-prone areas incorrectly predicted as non-flood-prone

▪

FP (False Positive) = number of non-ponding-prone areas incorrectly predicted as flood-prone

▪

TN (True Negative) = number of non-ponding-prone areas correctly predicted as non-flood-prone

ii.

SHAP Value Interpretability

SHAP values help to explain the spatial patterns seen in flood susceptibility maps by showing each feature’s impact on the output of the model [96,97]. They ensure model transparency and offer stakeholders evidence-based explanations for why certain regions are flagged as high or low-risk. This interpretability is vital for guiding infrastructure planning, flood mitigation efforts, and emergency response strategies [96,98,99]. SHAP values help to explain the spatial patterns seen in flood susceptibility maps by showing each feature’s impact on the output of the model [96,97]. They ensure model transparency and offer stakeholders evidence-based explanations for why certain regions are flagged as high or low-risk. This interpretability is vital for guiding infrastructure planning, flood mitigation efforts, and emergency response strategies [96,98,99].

iii.

Validation by High-Risk Locations

The high-flood-risk locations in Jefferson County are monitored by flood sensors, identified by local agency experts based on past flooding events and historic high-water marks [100]. In addition, the Local Community Task Force, which includes senior engineers and experts from the Jefferson County Drainage Districts, Texas Department of Transportation (TxDOT), U.S. Department of Energy (DOE), and other local agency experts, identified high flood-risk areas, many of which have historically been prone to flooding. Consequently, an effective model should be able to classify the majority of high-risk locations into high or very high flood-susceptibility zones.

iv.

Comparison with BLE Maps

Base Level Engineering (BLE) is a flood-mapping methodology designed to provide hazard information in ungauged or understudied areas. The BLE_DEP0_2PCT dataset quantifies estimated flood depths within the 1% annual chance floodplain by subtracting ground elevation from modeled water surface elevations, thereby generating raster-based flood depth values [101].

3. Results and Discussion

This section presents a comprehensive evaluation of the flood susceptibility modeling results. First, a correlation analysis is conducted to assess and address potential multicollinearity, followed by an evaluation of AUC scores to compare the predictive performance of different models. Then, SHAP values are analyzed to identify the most influential features. Following this, the generated flood susceptibility maps illustrate how each model delineates areas of potential flood risk. Finally, the models are validated against high-risk locations monitored by sensors, community task force flood-prone areas, and reference maps from BLE.

3.1. Correlation Analysis

As illustrated (Figure 8), the maximum and minimum correlation values among the continuous variables are 0.36 and −0.53, respectively. These results indicate that the continuous features are linearly independent and can therefore be used in the machine learning algorithms without concerns regarding multicollinearity. This independence helps ensure that the predictive models are not affected by high variance caused by highly correlated features and confirms that selected variables are proper for robust model development.

3.2. Statistical Model Results

The frequency ratio (FR) results show that flood susceptibility is strongly influenced by topography, distance to rivers, land use, and soil characteristics. Low elevations, low-gradient slopes, and areas within 100 m of rivers have high FR values, confirming they are the most flood-prone. Similarly, zones with high topographic wetness index (TWI), depressions, poorly draining soils (Group D), and agricultural or wetland land uses exhibit higher susceptibility, as they limit infiltration. In contrast, higher elevations, steep slopes, forested land, and areas far from rivers have FR < 1, indicating low flood risk (

Table 4. Frequency Ratio Calculations.

Name	Class	Total Area	Total Area Percentage	Number of Flood Events	Event Percentage	Frequency Ratio (Fr)	FR × 100	FR Weight
Dem	<0.55	543,130	20.132	11,812	26.339	1.308	130.834	130
	0.55–1.42	538,189	19.948	10,749	23.969	1.202	120.153	120
	1.42–3.52	538,873	19.974	8313	18.537	0.928	92.805	92
	3.52–7.12	538,586	19.963	5280	11.774	0.590	58.977	58
	>7.12	539,115	19.983	8692	19.382	0.970	96.993	96
Slope	<0.1	903,006	33.471	23,418	52.219	1.560	156.01	156
	0.1–1.6	1,111,196	41.188	13,145	29.311	0.712	71.166	71
	>1.6	683,691	25.342	8283	18.470	0.729	72.883	72
Distance from the roads	<120	523,825	19.416	5497	12.258	0.631	63.131	63
	120–414	554,926	20.569	6304	14.057	0.683	68.341	68
	414–921	540,024	20.017	7564	16.867	0.843	84.263	84
	921–2247	539,084	19.982	11,171	24.910	1.247	124.663	124
	>2247	540,034	20.017	14,310	31.909	1.594	159.411	159
Distance from Major Streams	<100	521,718	19.338	12,032	26.830	1.387	138.740	138
	100–150	181,893	6.742	3031	6.759	1.002	100.247	100
	150–200	163,546	6.062	2575	5.742	0.947	94.719	94
	200–250	149,130	5.528	2250	5.017	0.908	90.765	90
	>250	1,681,606	62.330	24,958	55.653	0.893	89.287	89
Distance from minor Streams	<25	271,804	10.075	4404	9.820	0.975	97.475	97
	25–50	198,580	7.361	3253	7.254	0.985	98.548	98
	50–75	253,717	9.404	4258	9.495	1.010	100.962	100
	75–100	191,096	7.083	3236	7.216	1.019	101.873	101
	>100	1,782,696	66.077	29,695	66.215	1.002	100.209	100
TWI	<4	558,867	20.715	6695	14.929	0.721	72.068	72
	4–5.1	510,613	18.926	5705	12.721	0.672	67.215	67
	5.1–7.5	548,889	20.345	7290	16.256	0.799	79.899	79
	7.5–10.7	532,726	19.746	9774	21.795	1.104	110.375	110
	>10.7	546,798	20.268	15,382	34.300	1.692	169.234	169
NDVI	<−0.135	593,018	21.981	10,045	22.399	1.019	101.902	101
	−0.13–0.21	486,977	18.050	6389	14.247	0.789	78.927	78
	0.21–0.40	543,431	20.143	7817	17.431	0.865	86.536	86
	0.40–0.53	534,539	19.813	12,035	26.836	1.354	135.44	135
	>0.53	539,928	20.013	8560	19.088	0.954	95.376	95
Soil Type	Clay	1,157,825	42.916	21,257	47.400	1.104	110.44	110
	Clay loam	292,744	10.851	4075	9.087	0.837	83.741	83
	Coarse sand	3381	0.125	39	0.087	0.694	69.394	69
	Fine sand	3380	0.125	139	0.310	2.474	247.400	247
	Loam	200,463	7.430	4265	9.510	1.280	127.99	127
	Sandy clay loam	68,769	2.549	1759	3.922	1.539	153.87	153
	Silt loam	76,439	2.833	1594	3.554	1.255	125.45	125
	Silty clay	334,235	12.389	3905	8.708	0.703	70.286	70
	Silty clay loam	98,628	3.656	2058	4.589	1.255	125.53	125
	Very fine sandy loam	9934	0.368	40	0.089	0.242	24.223	24
	water	61,629	2.284	811	1.808	0.792	79.166	79
	no-data	390,466	14.473	4904	10.935	0.756	75.556	75
Rock Unit	F S	27,520	1.020	322	0.718	0.704	70.390	70
	Qal	663,801	24.604	9705	21.641	0.880	87.955	87
	Qb	647,102	23.985	9958	22.205	0.926	92.576	92
	Qbb	79,950	2.963	2206	4.919	1.660	165.99	165
	Qbc	1,134,389	42.047	20,195	45.032	1.071	107.09	107
	Qbi	44,282	1.641	1271	2.834	1.727	172.67	172
	Qd	6689	0.248	46	0.103	0.414	41.371	41
	Ql	27	0.001	0	0.000	0.000	0.000	0
	Wa	94,133	3.489	1143	2.549	0.730	73.047	73
LULC	Agricultural Land	1,052,065	19.498	18,378	20.490	1.051	105.08	105
	Barren	7890	0.292	328	0.731	2.501	250.09	250
	Developed	457,827	16.970	4576	10.204	0.601	60.129	60
	Forested Upland	36,632	1.358	426	0.950	0.700	69.960	69
	Grassland and pasture	13,773	0.511	557	1.242	2.433	243.29	243
	Shrubland	4988	0.185	54	0.120	0.651	65.128	65
	Wetlands	998,782	37.021	17,493	39.007	1.054	105.36	105
	water	125,936	4.668	3034	6.765	1.449	144.93	144
Depression	<0	1,348,059	0.500	29,032	0.647	1.296	129.559	129
	>0	1,349,834	0.500	15,814	0.353	0.705	70.479	70
Average precipitation	<1554	535,490	19.848	10,057	22.426	1.130	112.984	112
	1554–1602	531,687	19.707	7978	17.790	0.903	90.269	90
	1602–1636	549,570	20.370	11,096	24.742	1.215	121.46	121
	1636–1695	542,972	20.126	10,274	22.910	1.138	113.83	113
	>1695	538,174	19.948	5441	12.133	0.608	60.821	60
Soil Hydrologic groups	A	31,459	1.166	741	1.652	1.417	141.701	141
	A/D	17,303	0.641	156	0.348	0.542	54.238	54
	B/D	95,411	3.537	2199	4.903	1.387	138.65	138
	C	1900	0.070	13	0.029	0.412	41.161	41
	C/D	78,197	2.898	1576	3.514	1.212	121.24	121
	D	2,473,623	91.687	40,161	89.553	0.977	97.672	97

).

The model highlights that flood-prone areas are concentrated in low-lying, flat, and poorly drained regions, near rivers and wetlands, while well-vegetated and elevated zones remain relatively safe.

3.3. Machine Learning Model

3.3.1. Hyperparameter Tuning

Hyperparameter optimization was conducted for both the Random Forest and XGBoost models to improve predictive performance. For the Random Forest, the parameter space included n_estimators, max_depth, min_samples_split, and min_samples_leaf. For the XGBoost model, the parameter space consisted of lambda, alpha, colsample_bytree, subsample, learning_rate, n_estimators, max_depth, and min_child_weight. The optimum hyperparameters obtained for both models are summarized in the

Table 5. Hyperparameters.

Random Forest		Xgboost
Parameter	Optimum Value	Parameter	Optimum Value
n_estimators	50	lambda	0.03251274199317688
max_depth	20	alpha	0.6440814745700857
min_samples_split	10	colsample_bytree	0.9
min_samples_leaf	4	subsample	0.6
		learning_rate	0.008
		n_estimators	1000
		max_depth	60
		min_child_weight	1

.

3.3.2. AUC Score

Among the models tested, XGBoost achieved the highest AUC of 0.92, confirming its superior ability to distinguish flood-prone from non-flood-prone areas. This aligns with its SHAP analysis, which revealed strong, targeted feature impacts for critical variables such as elevation, slope, and TWI. Additionally, XGBoost correctly classified the highest risk locations (100) in high or very high susceptibility zones, further validating its practical effectiveness. Random Forest, with an AUC of 0.88, also performed well and matched high-risk locations reasonably, though its SHAP values were more conservative. The CNN model had a lower AUC of 0.78, consistent with earlier findings that it underestimates risk, as it misclassified many flood-prone areas into low-susceptibility zones. Lastly, the Frequency Ratio model, with the lowest AUC of 0.65, lacked the complexity to model flood patterns accurately. These results clearly show that XGBoost offers the best balance of accuracy, interpretability, and real-world alignment, making it the most effective model for flood susceptibility prediction in Jefferson County (Figure 9).

3.3.3. SHAP Score

The SHAP value results from the CNN, Random Forest, and XGBoost models collectively offer strong insight into the role of various environmental and geospatial features in predicting flood susceptibility (Figure 10). Across all three models, elevation (DEM), Topographic Wetness Index (TWI), slope, NDVI, and distance to rivers or canals consistently emerge as the most influential features, confirming their critical roles in flood risk. Low elevation and flat terrain strongly increase flood susceptibility, as indicated by positive SHAP values, while higher elevations and steeper slopes have negative SHAP impacts, reducing predicted flood risk. TWI and NDVI further refine this by identifying areas prone to water accumulation or sparse vegetation, both conditions that heighten flood vulnerability. CNN and XGBoost tend to assign broader SHAP value ranges, meaning they respond more sharply to high-risk conditions compared to the Random Forest model, which produces more conservative, narrowly distributed SHAP values [102]. XGBoost emphasizes extreme flood-prone areas more than safe zones, evident in its asymmetric SHAP distribution, whereas CNN shows a more balanced response across the risk spectrum. In contrast, the Random Forest model shows many features with zero SHAP values, particularly for less impactful variables like soil type, rock unit, land cover (NLCD), and long-term precipitation, indicating they often do not affect predictions in that model. However, the models still agree that these features are less critical overall.

3.4. Flood Susceptibility Maps

Four different flood susceptibility maps were delineated using machine learning and flood frequency weights data in ArcGIS Pro. The flood frequency weight model as a statistical model, results reveal significant variations in flood susceptibility based on topographic, hydrologic, and environmental factors. Dem (<0.55), Lower slope areas (<0.1°) and regions closer to major streams (<100 m) exhibit notably higher flood susceptibility. Additionally, higher Topographic Wetness Index (TWI > 10.7) and specific land-use categories such as barren lands and grasslands have markedly elevated susceptibility. Soil type plays a critical role, with fine sand and sandy clay loam demonstrating notably high weights. Hydrologic groups A and B/D also indicate increased flood occurrence. To help visualize the results, the continuous flood risk values were grouped into five classes: very low, low, moderate, high, and very high susceptibility, using a quantile classification method. This method helps show changes across the area more clearly and avoids using fixed intervals. The maps showed similar patterns in areas marked as very low and low risk, especially in higher or sloped land. But there were noticeable differences in areas with moderate to very high risk, mostly in flat, low-lying regions with water pixel concentration in their target. These high-risk areas were mainly found in the central and southern parts of Jefferson County (Figure 11).

3.5. High Flood-Risk Locations Monitored by Flood Sensors

Figure 12 illustrates the effectiveness of each model, including Frequency Weight, Random Forest, XGBoost, and CNN, in capturing flood-prone areas based on the high-flood-risk locations. Looking at the results, the XGBoost model shows the best alignment with high-risk locations: 39 high-risk locations fall in the “High Susceptibility” category and 61 in the “Very High Susceptibility” category, totaling 100 high-risk locations (out of 121). This is the highest concentration of sensors in flood-prone zones among all models, indicating that XGBoost can map the areas with the highest perceived flood risk. The Random Forest model is slightly behind, with 66 high-risk locations classified as high or very high susceptibility. CNN, on the other hand, only places 50 high-risk locations in these top two risk categories, suggesting it may underpredict flood hazards in historically vulnerable areas.

Two locations were selected for the study, both recognized as high-risk areas monitored by flood sensors BM27 and BM12. Case 1, associated with flood sensor BM27, is located in a residential area that includes a ditch designed to direct excess floodwater along the street. Case 2, linked to flood sensor BM12, represents another high-risk location that is fully developed with impervious concrete pavement and streets, also with a ditch to direct excess floodwater. In Figure 13, in the first row, all three models, CNN, RF, and XGBoost, successfully identify a high-risk location monitored by sensor BM27 within a very high flood susceptibility zone (red area), demonstrating basic model competence. However, the XGBoost model outperforms the others by identifying neighboring locations as high-risk areas that closely align with water pixel patterns shown in blue. While CNN and RF models produce dispersed predictions around the high-risk locations, XGBoost captures a broader, more realistic spatial extent of vulnerability. In the second row, for sensor BM12, the performance of XGBoost becomes even clearer, as it maps a concentrated and well-defined high-risk zone around the sensor, in strong agreement with water pixel concentration. CNN and RF fail to provide the same spatial consistency, offering scattered predictions that miss portions of the flooded region. Overall, XGBoost demonstrates better accuracy and reliability in both detecting known flood-prone zones and representing the surrounding areas at risk.

3.6. Comparison with the BLE Map

Compared with the Base Level Engineering (BLE), BLE_DEP0_2PCT also shows a good agreement with the XGBoost Flood susceptibility map. Although some areas were not detected, it’s because the XGBoost model is based on pluvial information, while BLE_DEP0_2PCT is a raster dataset class that contains the estimated flood depth throughout the 1% annual chance floodplain determined during the Base Level Engineering assessment. This dataset is compiled by performing a calculation to remove the ground elevation from the water surface elevation dataset, thereby calculating the depth of flooding expected within the 1% annual chance floodplain extents (Figure 14).

3.7. High-Risk Areas Identified by the Local Community Taskforce

All of these areas were also classified as highly susceptible by the XGBoost model (Figure 15).

4. Conclusions

The study developed a flood inventory using multi-year Sentinel-1 SAR data (2018–2023). Based on this inventory, a Frequency Ratio analysis and three machine learning models, Random Forest, CNN, and XGBoost, were then applied to generate flood susceptibility maps, identifying high-frequency ponding hot spots across Jefferson County. The maps were validated against known high-risk areas, the BLE risk map, and the community task force.

• The alignment of the XGBoost model with the BLE risk map and the community task force assessments indicates strong qualitative agreement and model reliability.
• The XGBoost model achieved the best performance (AUC = 0.92), correctly classifying 100 of 121 sensor-identified high-risk locations, outperforming all other methods. Random Forest showed good accuracy (AUC = 0.88) but underestimated severity, while CNN (AUC = 0.78) misclassified many high-risk areas. The Frequency Ratio model had the weakest predictive power (AUC = 0.65), confirming XGBoost as the most reliable approach for flood susceptibility mapping in Jefferson County.
• SHAP value analysis further validated the XGBoost model interpretability, revealing that elevation, slope, TWI, and NDVI were consistently the most influential features. This helps researchers, engineers, and policymakers by highlighting the key environmental factors that drive flood susceptibility.
• Policymakers can use flood susceptibility maps to identify high-risk hotspots and develop optimized early warning systems. By integrating these maps with real-time rainfall and sensor data, flood-prone areas can be predicted more accurately. This supports timely alerts, efficient resource allocation, and improved evacuation planning, ultimately enhancing community preparedness and reducing losses.

This study utilized six years of Sentinel-1 SAR-derived water pixel data to model flood susceptibility; however, the lack of comprehensive real flood event records may have limited the results’ precision. Using finer-resolution DEM and radar imagery could improve model accuracy but would significantly increase computation time. Further optimization of VV and VH polarization thresholds for the area of study and the inclusion of a hydrological inundation model can enhance classification around major and minor river networks. Future research should refine geospatial factors to improve susceptibility accuracy and test the methodology in mountainous regions, with the related variables, which could strengthen the reliability and generalizability of the approach.