Modeling Flood Susceptibility in Rwanda Using an AI-Enabled Risk Mapping Tool

Abstract

This study presents the development of a Python-based flood-susceptibility risk-mapping tool, implemented in Jupyter Notebook, applied to Rwanda. A Flood Susceptibility Index (FSI) was developed by integrating 20 causal factors associated with flood occurrences, including topographic, hydrological, geological, and anthropogenic variables. Logistic regression, and Variance Inflation Factor were implemented in Python using libraries such as Numpy, Arcpy, traceback, scipy, Pandas, Seaborn, and statsmodel to assign weights to each factor, and to address multicollinearity. The model was validated against flood extent data derived from Sentinel-1 satellite imagery for the major historical flood event that occurred from 2014 to 2024, ensuring spatial consistency and predictive reliability. To project future flood susceptibility for 2030, precipitation data from the Institut Pierre Simon Laplace Coupled Model, version 5A, Medium Resolution (IPSL-CM5A-MR) climate model under the Representative Concentration Pathway 8.5 (RCP 8.5) scenario were utilized. The resulting FSI was classified into five susceptibility levels, from very low to very high, and visualized using Python’s geospatial and plotting tools within Jupyter Notebook in ArcGIS Pro 3.5. It indicates that areas with high amounts of rainfall, and proximity to wetlands and rivers reveal the highest flood risk. The automated and reproducible approach offered by Python enhances transparency and scalability, providing a decision-support tool for disaster risk reduction and climate adaptation planning in Rwanda.

1. Introduction

Floods are among the most frequent and devastating natural hazards, posing a serious threat to human life and causing substantial economic losses worldwide [1]. It is estimated that more than one-third of the global land area is susceptible to flooding [2,3]. The intensity and frequency of flood events have been exacerbated by anthropogenic factors, including land-use changes and climate change, which have altered hydrological cycles and increased runoff [4,5]. Floods and landslides are particularly unavoidable and are expected to become more dangerous in the future, impacting numerous regions around the world. In Africa, floods are a major natural hazard, having caused over 27,000 fatalities between 1950 and 2019, and affecting millions of people through displacement, infrastructure damage, and widespread economic losses [6,7]. Unlike other continents, floods in Africa are predominantly triggered by intense rainfall events. Consequently, countries experiencing extreme precipitation are especially vulnerable to frequent and severe flood incidents [8,9].

According to the Ministry in Charge of Emergency Management (MINEMA) of Rwanda, floods are ranked as the second most severe disaster in the country after landslides. In 2021 alone, floods resulted in numerous fatalities and were the leading cause of crop damage across large areas [10]. Given the current trends and projected future flood scenarios, there is a growing need for comprehensive spatial and temporal information on flood-prone areas. In particular, regions characterized by flat riverine topography, combined with rapid and unplanned urbanization and inadequate drainage infrastructure, face exacerbated flood risks in river basins and deltaic areas [11,12].

Flood susceptibility mapping (FSM) plays a vital role in disaster risk management by identifying areas with different levels of flood likelihood based on environmental and geological factors. While traditional FSM relied on empirical and deterministic methods, modern approaches increasingly favor data-driven and statistical models such as logistic regression [13,14]. These models are preferred for their objectivity, simplicity in interpretation, and ability to efficiently analyze multiple contributing factors simultaneously. Logistic regression model, in particular, estimates the probability of flood occurrence using independent variables such as slope, rainfall, land cover, and soil characteristics [15,16].

Recent advancements in open-source programming and geospatial technologies, particularly in Python have enabled the development of automated and scalable flood susceptibility mapping (FSM) workflows. Python offers a flexible and robust platform for geospatial data processing, statistical modeling, and visualization through powerful libraries [17,18].

This integration is particularly relevant in resource-constrained contexts like Rwanda, where accessible and reproducible tools can support national agencies in proactive hazard monitoring and planning.

Most previous studies on Flood Susceptibility Index (FSI) mapping have relied on traditional statistical or multi-criteria decision-making methods such as the Analytical Hierarchy Process (AHP), Frequency Ratio (FR), and Weight of Evidence, which often depend heavily on expert judgment, linear assumptions, or predefined weighting schemes [19,20]. While these approaches have provided valuable baseline assessments, they may not fully capture the complex, nonlinear interactions among hydrological, geomorphological, climatic, and anthropogenic factors that drive flood occurrence. The key knowledge gap this study intends to address lies in the limited ability of conventional methods to model high-dimensional, nonlinear relationships and dynamically learn from large, multi-source datasets. By applying a Machine Learning Based Logistic Regression approach, this study aims to enhance predictive accuracy, reduce subjectivity in parameter weighting, improve generalization capability across diverse landscapes, and provide a more robust, data-driven flood susceptibility framework that better supports risk management and decision-making under changing climate conditions.

Rwanda Space Agency [21] developed a Python-based logistic regression model to produce a Flood Susceptibility Index (FSI) map to select the high-risk regions in Rwanda. The objectives are to (1) demonstrate a replicable, open-source methodology for Flood Susceptibility mapping, (2) assess the model’s predictive performance using standard statistical metrics, including Pearson correlation and the Receiver operating characteristic (ROC) curve and the Area under the Receiver Operating Curve (AUC) or c-statistic to assess how well a Logistic Regression Model predicts flood susceptibility, (3) provide an actionable FSI map to aid national disaster management and land-use planning, and (4) predict the flood susceptibility using precipitation projections dataset from the IPSL-CM5A General Circulation Model (GCM) under the Representative Concentration Pathway (RCP) 8.5 scenario, as part of the Coupled Model Intercomparison Project Phase 5 (CMIP5).

2. Study Area and Materials

2.1. Description of the Study Area

Rwanda is a hilly and mountainous country, with elevations ranging from 913 m to 4431 m above sea level (Figure 1b), and it enjoys a tropical temperate climate largely influenced by its high altitude [22,23]. The country covers a surface area of approximately 26,338 km² and is situated on the eastern flank of the Kivu-Tanganyika Rift in the East African Rift System (Figure 1a) [24]. Although landlocked, Rwanda possesses abundant water resources (Figure 1b), including lakes that cover about 128,190 hectares, wetlands spanning 77,000 hectares, and an estimated 7260 hectares of watercourses. The country experiences two main rainy seasons: the first extends from March to May, and the second occurs from October to November, with average monthly rainfall ranging between 110 mm and 200 mm [25,26]. In contrast, the dry seasons are divided into a short period from December to February and a longer period from June to early September [27].

Administratively, Rwanda is divided into five provinces: The Northern, Southern, Eastern, and Western Provinces, along with Kigali City, the capital. Each province is further subdivided into five to eight districts, forming the second level of local government [28].

2.2. Factors Influencing Floods in Rwanda

A wide range of environmental and anthropogenic factors influences floods. However, identifying a universally accepted set of parameters for flood susceptibility mapping remains a challenge due to the complexity and variability of flood dynamics. Understanding the underlying factors that contribute to flooding in a specific region is crucial for selecting the most relevant parameters in predicting and managing future flood occurrences [29,30]. In this study, a total of 20 flood causative factors were selected based on logistic regression model analysis [31]

These factors were grouped into topographic, climatic, and Others, as described below:

2.2.1. Topographic Features

Digital Elevation Model (DEM)

Topographic parameters were derived from a 10 m spatial resolution DEM obtained from the Rwanda Housing Authority (RHA) [32]

Topographic Ruggedness Index (TRI)

The topographic ruggedness (Roughness) index (Equation (1)) was developed by Riley et al. (1999), which measures the elevation difference between a center cell and its eight surrounding cells, providing an estimate of terrain variability [33,34].

$$TRI=\sqrt{Abs(E_{max}^2-E_{min}^2)}$$

(1)

where E_max and E_min represent the largest and smallest values of cells in nine rectangular neighborhoods of altitude values.

Topographic Wetness Index (TWI)

TWI (Equation (2)) reflects the influence of topography on surface runoff and potential water accumulation at any given point in a watershed [34,35].

$$TWI=ln\left({\displaystyle \frac{A_s}{tan\beta }}\right)$$

(2)

where A_s is the catchment area or flow accumulation (m²/m) and β is the local slope gradient measured in degrees.

Slope Percent (%)

Slope affects runoff speed and infiltration rates. Flooding tends to occur more readily in low-slope areas due to slower runoff velocities, while steep slopes encourage rapid drainage and reduced flood accumulation [36,37]. Therefore, the slope percent was generated using the ArcGIS focal statistics tool [38].

2.2.2. Climate Features

Climate data layers, including precipitation, surface temperature, and the Antecedent Precipitation Index (API) were derived from the Enhancing National Climate Services (ENACTS) initiative by Columbia University’s International Research Institute for Climate and Society (IRI). These datasets, developed at ~4 km resolution (0.0375°), combine ground station data from the Rwanda Meteorology Agency (Meteo Rwanda) with satellite data from Tropical Applications of Meteorology using SATellite and ground-based observations (TAMSAT). Temperature estimates incorporate JRA-55 reanalysis data and national observations [39].

Precipitation

The precipitation is the most critical factor influencing flooding in Rwanda, with flood frequency increasing in areas of lower elevation and higher rainfall [3,29].

Observed Precipitation Data

Observed monthly precipitation data (Figure 2) were obtained from the Rwanda Meteorology Agency, which covers over 36 years (1981–2017). Based on experiences, research, and reports on flood incidences in Rwanda [40], heavy precipitation/rainfall is the most causal factor of floods in Rwanda. Nevertheless, the precipitation was taken into account in the production of the Flood Susceptibility Index.

Predicted Annual Precipitation Data

The projected precipitation data from Institut Pierre Simon Laplace (IPSL) originated from the Coupled Model Intercomparison Project—Phase 5 (CMIP5) and was chosen based on Pearson correlation with in situ data from Meteo Rwanda and RCP 8.5 using 60 weather stations. This present study used ipsl_cm5a_mr_rcp8_5_2030s_prec_2_5min_r1i1p1_an as the best RCP with a correlation of 0.627 (Appendix A,

Table A2. The top 10 best-performing RCP models using Pearson correlation with in situ data from 60 stations of Meteo Rwanda.

RCP Data	Pearson Correlation	ID
ipsl_cm5a_mr_rcp8_5_2030s_prec_2_5min_r1i1p1_an	0.6267	p131
ipsl_cm5a_mr_rcp2_6_2030s_prec_2_5min_r1i1p1_an	0.6243	p127
ipsl_cm5a_lr_rcp8_5_2030s_prec_2_5min_r1i1p1_an	0.6239	p125
ipsl_cm5a_mr_rcp4_5_2030s_prec_2_5min_r1i1p1_an	0.6236	p129
ipsl_cm5a_lr_rcp4_5_2030s_prec_2_5min_r1i1p1_an	0.6222	p121
gfdl_esm2m_rcp4_5_2030s_prec_2_5min_r1i1p1_an	0.6203	p091
ipsl_cm5a_lr_rcp6_0_2030s_prec_2_5min_r1i1p1_an	0.6193	p123
ncar_ccsm4_rcp8_5_2030s_prec_2_5min_r1i1p1_an	0.6186	p201
csiro_access1_3_rcp8_5_2030s_prec_2_5min_r1i1p1_an	0.6182	p053
cesm1_cam5_rcp4_5_2030s_prec_2_5min_r1i1p1_an	0.6174	p035

). Global General Circulation model IPSL-CM5a was developed to study the long-term response of the climate system to natural and anthropogenic forcing as part of the 5th Phase of the Coupled Model Intercomparison Project (CMIP5) [41]

Antecedent Precipitation Index (API)

The API was computed based on the empirical (Equation (3)) [42]. The total five-day API was calculated by summing the daily API [43], estimated from gridded daily ENACT precipitation (~4 km resolution). Note that this parameter is used because reliable data on soil moisture is unavailable.

$${\mathrm{I}}_{\mathrm{t}}=\mathrm{k} {\mathrm{I}}_{\mathrm{t}-1}+{\mathrm{P}}_{\mathrm{t}}$$

(3)

where it is the API estimated at the end of each day _t; P_t is the precipitation amount during day _t; and k is a decay (recession) factor representing a logarithmic decrease in soil moisture with time during periods of no precipitation. This study used a recommended k-factor value of 0.9 [44,45].

Surface Temperature

Surface temperature, particularly Land Surface Temperature (LST), plays a significant role in influencing flood occurrences by affecting evapotranspiration, soil moisture retention, and rainfall patterns [46]. High surface temperatures often caused by urbanization, deforestation, and climate change, can lead to the drying and hardening of soil surfaces, reducing their capacity to absorb rainfall [47].

When intense precipitation occurs after prolonged periods of heat, reduced infiltration leads to increased surface runoff, thereby elevating the risk of flooding [48]. Thus, incorporating surface temperature data into flood susceptibility modeling helps capture the indirect yet critical pathways through which thermal characteristics of the land influence hydrological extremes.

2.2.3. Others (Land Use and Geological Features)

Land Use/Land Cover (LULC)

LULC data were obtained from a 10 m resolution ESA World Cover dataset (2021), derived from Sentinel-2 imagery [49].

Lithology

Lithological soil data at 250 m resolution were sourced from the Soil and Terrain Database for Central Africa (SOTERCAF) [50].

Soil Texture

The soil layers of clay, silt, and sand fractions for Rwanda from the Africa Soil Information Service (AfSIS) at 250 m spatial resolution [50] were used to derive the soil texture map of Rwanda based on the soil texture triangle classification system developed by the United States Department of Agriculture (USDA). It is built considering its percentage of clay, silt, and sand, ranging from the fine textures (clay) to the intermediate textures (loam) and the coarser textures (sand) [51].

Roads

The road network layer was generated from 2021 vector datasets provided by the World Food Programme (WFP) using Euclidean distance analysis in ArcGIS at 10 m resolution.

Faults and Earthquake Hotspots (1950–2022)

Data on fault lines and seismic activity from 1950 to 2022 were acquired from the Rwanda Mines, Petroleum, and Gas Board (RMB) [52]. A proximity layer to faults was created using Euclidean distance analysis.

Drainage Density

Drainage density, an indicator of how water drains across the surface (Equation (4)), was calculated in ArcGIS Pro using the Spatial Analyst toolbox [53,54].

$$D_d={\displaystyle \frac{\sum \mathrm{L}}{A_{basin}}}$$

(4)

where D_d is the drainage density, ∑L: total length of draining channels in a basin, A is the surface area of a basin.

2.3. Source of Data

The study used both raster and vector datasets derived from open and institutional sources, covering the main physical and environmental factors known to influence flood occurrence according to the literature review (

Table 1. Data sources of the data used in this study.

Factor Type	Dataset	Resolution	Source
Topography	Mean elevation, Max. altitude within 4 × 4 matrix, Mean slope %, Max. slope % within 4 × 4 matrix, Slope aspect, Topographic ruggedness index (TRI), Topographic wetness index (TWI) or compound topographic index (CTI), Slope length (LS), Flow accumulation, Stream power index (SPI), Topographic curvature, Topographic plan curvature, Topographic profile curvature, Drainage density.	10 m	Rwanda Housing Authority
Climate	Long-term daily antecedent precipitation index (API), Long-term daily precipitation, Long-term annual precipitation, Long-term daily and annual mean surface temperature	~4 km	ENACTS products from Meteo Rwanda [39]
Soil	Soil texture, Soil permeability/erodibirity (K-factor)	250 m	AfSIS’s Soil properties V1 [50]
	For soil lithology & texture, their parameters were used: CFRAG: coarse fragments (>2 mm); SDTO: sand (mass %); STPC: silt (mass %); CLPC: clay (mass %); BULK: Single bulk density (kg dm−3); TAWC: available water capacity (cm3 cm−3 102, −33 kPa to −1.5 MPa conform to USDA standards); CECs: Single cation exchange capacity (cmol c kg−1) for fine earth fraction; BSAT: base saturation as percentage of CECsoil; CECc: Single CECclay, corrected for contribution of organic matter (cmol c kg−1); PHAQ Single pH measured in water; TOTC Single organic carbon content (g C kg−1); TOTN: Single total nitrogen (g N kg−1); CNrt: Single C/N ratio; ECEC Single effective CEC (cmol c kg−1).	-	SOTERCAF v.1.0 [55], https://files.isric.org/public/sotwis/SOTWIS_CAF.zip (accessed on 17 May 2024)
	Soil depth	3 classes	MINEMA-RAB
Land cover	Land cover classification	10 m	ESA Sentinel-2 (https://viewer.esa-worldcover.org/worldcover/, accessed on 21 September 2023)
	Normalized Difference Vegetation Index (NDVI) Normalized Difference Water Index (NDWI), Transformed Normalized Difference Vegetation Index (TNDVI)	10 m	ESA Sentinel-2 (https://dataspace.copernicus.eu/ accessed on 21 September 2025)
Faults	Distance from faults and earthquake hotspots (1950–2022)	Lines	(https://www.rmb.gov.rw/ accessed on 21 September 2025)
Roads	Distance from road, Roads and streets	Lines	World Food Programme (WFP)
Rivers	Distance from rivers and streams	Lines	Rwanda Housing Authority

).

2.4. Data Processing

All spatial data were first projected to a common coordinate system (WGS 84/UTM Zone 35S) and resampled to a unified 10 m spatial resolution, the original spatial resolution of RHA DEM. The preprocessing steps included: Raster layers were converted to Geodatabases, Euclidean distance tool was used to generate raster layers of the distance from faults, rivers and roads at 10 m resolution and converted them into Geodatabases and a layer of flood incidents Inventory in points was converted in Geodatabase.

Classified layers (i.e., land cover classification, soil texture, soil lithology and depths) were converted into Geodatabases after converting them into raster layers with continuous values based on the distribution or percentage (%) of total flood incidences in their different classes (

Table 2. Scores for classifiable layers (Land Cover, Lithology, Soil texture, and Soil Depth).

Land Cover	Flood Sample (FSI)_Truth	FS_Factor (%)
Wetland	30,127	75.3
Forest	3483	8.7
Shrubs	198	0.5
Grassland	1614	4
Cropland	3901	9.8
Urban	177	0.4
Bare land	203	0.5
Waterbodies	295	0.7
Lithology	FS_truth	FS_factor
Volcanic ash	637	1.6
Colluvial	518	1.3
Schist	11,625	29.1
Granite	150	0.4
Quartzite	37	0.1
Fluvial	11,109	27.8
Basic igneous rock	822	2.1
Organic	14,431	36.1
Water	552	1.4
Acid metamorphic rock	0	0
Basalt	98	0.2
Lacustrine	0	0
Gneiss, migmatite	0	0
Soil texture	FS_truth	FS_scores (%)
Loam	141	0.4
Silty loam	0	0
Sandy clay loam	6246	15.6
Clay loam	26,364	66
Silty clay loam	18	0
Sandy clay	1199	3
Silty clay	7	0
Clay	5989	15
Soil Depth	FS_truth	FS_scores (%)
<0.5 m	13,233	33.1
0.5–1 m	24,031	60.1
>1 m	2734	6.8

).

3. Methodology

3.1. Methodology Overview and Workflow

This study applies a logistic regression model within a Python 3.9 environment to generate a Flood Susceptibility Index (FSI) map using a combination of flood-triggering factors. The methodology includes five main stages: data collection, preprocessing, variable selection, model development, and FSI map generation (Figure 3). The entire process was implemented using open-source Python libraries, enabling reproducibility and scalability. The Flood Susceptibility Index (FSI) map of Rwanda was developed using a data-driven logistic function model adapted from the logistic function model used to develop the map of flood Susceptibility Index of Africa and Flood Prediction in Africa [56,57]. The regression coefficients between flood incidence triggering factors were calculated following the study on runoff prediction over the African continent [58]. The logistic model is a statistical model that models the log-odds of an event as a linear combination of one or more independent variables. In regression modeling, logistic regression estimates the coefficients in the linear or nonlinear combinations. In binary logistic regression, there is a single binary dependent variable, coded by an indicator variable, where the two values are labeled “0” and “1”, while the independent variables can each be a binary variable (two classes, coded by an indicator variable or a continuous variable [59].

The flexibility of field samples integration in logistic model make it robust than other methods such as the Analytical Hierarchy Process (AHP) that is widely used for FSI and FSI modeling [60,61,62,63]. AHP relies heavily on expert judgments for pairwise comparisons, which can introduce subjective biases. Different experts might have different opinions, leading to inconsistency in results. AHP can be sensitive to the scaling method used for judgments. Differences in interpretation of scales (e.g., 1 to 9 scale) can significantly impact outcomes. AHP is time-intensive for large-scale problems. The process requires time and commitment to complete all pairwise comparisons, making it less practical for large, time-sensitive decisions [64,65].

3.2. Collinearity Analysis of Influencing Variables

To ensure the statistical robustness of the flood susceptibility model, a collinearity assessment was conducted on the selected topographic predictor variables. Collinearity, or multicollinearity, occurs when two or more independent variables in a model are strongly correlated with one another, which can inflate standard errors, reduce coefficient stability, and ultimately compromise the reliability and interpretability of model predictions [66,67]. Topographic variables were carefully assessed for redundancy and were either excluded or substituted to improve model parsimony. This step was essential to enhance the stability and interpretability of the model used in the subsequent analysis. To further address potential multicollinearity, the Variance Inflation Factor (VIF) was calculated for the topographic features (Appendix A,

Table A5. The selected topographic factors based on multicollinearity and VIF.

DEM Factors	VIF
Altitude	3.162016
TWI	3.010847
Drainage	1.645533
Aspect	1.162369
TPI	1.129553
Faccum	1.005082

). A VIF value greater than 10 indicates a high level of multicollinearity, which may compromise the model’s predictive performance and reliability [68]

Figure 3 illustrates the workflow used for modeling the Flood susceptibility index (FSI) for Rwanda.

3.3. Workflow of Development of the Flood Susceptibility Index

Code available on: https://github.com/Valentinemu/flood-susceptibility-index.git (accessed on 13 February 2026).

3.4. Logistic Regression Model Development

3.4.1. Logistic Regression Theory (Brief Overview)

Logistic regression is a statistical classification method used to model the probability of a binary outcome based on one or more independent variables. In the context of flood susceptibility, it is used to predict the likelihood that a flood will occur (1) or not occur (0) based on factors such as slope, rainfall, soil type, and land cover, etc.

Unlike linear regression, which predicts continuous values, logistic regression models the log-odds of the dependent variable as a linear combination of the input variables (Equation (5)):

$$\mathit{log}\left({\displaystyle \frac{P}{1-P}}\right)={\beta }_0+{\beta }_1{X}_1+{\beta }_2{X}_2+\cdots +{\beta }_n{X}_n$$

(5)

where:

P is the probability of a flood occurrence;

X₁, X₂, …, X_n are the independent (triggering) variables;

β₁, …, β_n are the coefficients (weights) associated with each variable.

This can be rewritten to express the probability as:

$$P={\displaystyle \frac{1}{1+e^{-({\beta }_0+{\beta }_1{X}_1+\cdots +{\beta }_n{X}_n)}}}$$

(6)

After training the model using a known dataset of flood and non-flood locations, the coefficients β are estimated using maximum likelihood estimation (Equation (6)). The resulting model can then be applied to new spatial data to generate a flood Susceptibility Index (FSI) map, where each location is assigned a probability of flood occurrence. This method was summarized based on the GIS-based logistic regression for Flood susceptibility mapping [69,70,71,72].

3.4.2. Logistic Regression Implementation in This Study

Variable Selection

The 31 flood triggering factors were evaluated using 39,998 flood points and 17,461 points of non-flood locations. The Pearson correlation (r) and p-values were used to test the statistical significance of each variable. Only features with a highly significant p-value < 0.00 [73] were selected for FSI modeling.

To avoid multicollinearity among predictor variables, a correlation matrix and p-value analysis were performed using the statsmodel 0.14.0 library. Variables with a high correlation p-value > 0.001 were excluded to improve model stability [73]. This analysis indicated that 20 features are major flood triggering factors (p-value < 0.00) (

Table 3. The selected factors after applying the Pearson correlation and VIF.

Factors	Pearson Correlation (Flood Incidence and FSI Factor)	p Value	Number of Samples (Flood and Non-Flood Points)
Mean Altitude	0.657	0.00	57,459
Antecedent Precipitation Index (API)_mm_daily_ENACTS1991_2020	0.549	0.00	57,459
Distance from Faults and earthquake hotspots (1950–2022)	0.495	0.00	57,459
Distance_from_road	0.045	0.00	57,459
Landcover_factor_from_flood_incidences_count_per_landcover_type	0.701	0.00	57,459
Lithology_factor_from_flood_incidences_count_per_lithology_type	0.499	0.00	57,459
Precipitation_mm_annual_ENACTS1983_2017	0.627	0.00	57,459
Soil_depth_factor_from_flood_incidences_count_per_soildepth_range	0.355	0.00	57,459
Soil texture_factor_from_flood_incidences_count_per_soiltexture_type	0.050	0.00	57,459
Surface_temperature_oC_mean_daily_ENACTS1991_2020	0.571	0.00	57,459
Topographic wetness index (TWI) or compound topographic index (CTI)	0.576	0.00	57,459
TPI (Topographic Position Index)	0.069	0.00	57,459
Drainage density	0.317	0.00	57,459
NDVI	0.360	0.00	57,459
Aspect	0.160	0.00	57,459
Soil_permeability_erodibirity_K_factor	0.233	0.00	57,459
Slope length (LS)	0.046	0.00	57,459
Flow accumulation	0.031	0.00	57,459
Distance from the river	0.018	0.00	57,459
Normalized Difference Water Index (NDWI)	0.385	0.00	57,459

). Figure 4 illustrates the correlation heat map of the influencing factors of flood.

The present study implemented the logistic regression model using scikit-learn, referring to the published study on FSI assessment in Africa [56]. The regression coefficients c3, c2, c1, and b were calculated with the python’s linregress function. In the present study, the regression coefficients c3, c2, c1, and b were calculated based on the training samples of 39,998 flood points and 17,461 non-flood points (stable locations) and their corresponding values of significant flood features.

The response variable was binary:1 for locations with known flood occurrence, and 0 for randomly sampled non-flood locations.

Model Equation:

$${\mathrm{F}\mathrm{S}\mathrm{I}}_{\mathrm{F}}={\displaystyle \frac{1}{1+\mathrm{E}\mathrm{x}\mathrm{p}\left(-\left({\mathrm{c}}_3\ast {\mathrm{x}}^3+{\mathrm{c}}_2\ast {\mathrm{x}}^2+{\mathrm{c}}_1\ast \mathrm{x}+\mathrm{b}\right)\right)}}$$

(7)

$$F_i=\left({\mathrm{c}}_3\ast {{FSI}_{Fi}}^3\right)+\left({\mathrm{c}}_2\ast {{FSI}_{Fi}}^2\right)+\left({\mathrm{c}}_1\ast {FSI}_{Fi}\right)+\mathrm{b}$$

(8)

$$\begin{gathered} {\mathrm{c}}_3=\mathrm{I}\mathrm{N}\mathrm{D}\mathrm{E}\mathrm{X}\left(\mathrm{L}\mathrm{I}\mathrm{N}\mathrm{E}\mathrm{S}\mathrm{T}\right(F_i,{FSI}_{Fi}^\{1,2,3\}),1) \\ {\mathrm{c}}_2=\mathrm{I}\mathrm{N}\mathrm{D}\mathrm{E}\mathrm{X}\left(\mathrm{L}\mathrm{I}\mathrm{N}\mathrm{E}\mathrm{S}\mathrm{T}\right(F_i,{FSI}_{Fi}^\{1,2\}),1,2) \\ {\mathrm{c}}_1=\mathrm{I}\mathrm{N}\mathrm{D}\mathrm{E}\mathrm{X}\left(\mathrm{L}\mathrm{I}\mathrm{N}\mathrm{E}\mathrm{S}\mathrm{T}\right(F_i,{FSI}_{Fi}^\{1,2\}),1,3) \\ \mathrm{b}=\mathrm{I}\mathrm{N}\mathrm{D}\mathrm{E}\mathrm{X}\left(\mathrm{L}\mathrm{I}\mathrm{N}\mathrm{E}\mathrm{S}\mathrm{T}\right(F_i,{FSI}_{Fi}^\{1,2\}),1,4) \end{gathered}$$

where FSI_F is a significant flood feature or factor adjusted or fitted with observed flood incidences and non-flood incidences (Equation (7)), x³, x², and x are the dependent variables (unadjusted flood factors) and c3, c2, c1, and b are the regression coefficients obtained based on training samples of flood points and non-flood (stable locations); Fi: samples of flood incidences (points) (1) and non-flood points (0) (Equation (7)), FSI: Values of each significant flood factor at the location of flood incidences and non-flood points.

For those who are not familiar with programming, can calculate the regression coefficients using other tools such as Microsoft (MS) Excel’s built-in LINEST array function [58]. In addition, the symbol * in equations represents multiplication.

Data split into training (70%) and testing (30%) sets. Model fitted using logistic regression from sklearn linear model. Model evaluated using Accuracy, Confusion Matrix, Receiver Operating Characteristic (ROC) curve, and Area Under the Curve (AUC).

Sentinel-1 satellite imagery was used to detect historical flood events that occurred across Rwanda between 2014 and 2024, as illustrated in Figure 5. The analysis identified all flooded areas nationwide during this period, capturing both the extent and spatial distribution of events.

Flood occurrences were mapped using consistent radar-based detection to ensure accuracy and comparability over time. Each flood event was georeferenced to its precise location to support spatial indexing. Locations that experienced multiple floods were assigned higher event counts. This indexing approach highlights areas with repeated flood exposure. It provides a clear understanding of Rwanda’s flood-prone zones over the past decade.

3.5. FSI Map Generation

Generation of the National Flood Susceptibility Index (FSI) map of Rwanda: Using Equation (9), the FSI map of Rwanda was obtained through a combination of all 20 layers of significant flood factors derived with a logistic model and multiplication of their correlation (weight) to each factor layer (Pearson correlation).

$$\begin{array}{c}\mathrm{F}\mathrm{S}\mathrm{I}={\displaystyle \frac{{\left[\right({\mathrm{F}\mathrm{S}\mathrm{I}}_{\mathrm{F}\mathrm{t}\mathrm{o}\mathrm{p}}}_1\ast {\mathrm{r}}_1)+\cdots ({{\mathrm{F}\mathrm{S}\mathrm{I}}_{\mathrm{F}\mathrm{t}\mathrm{o}\mathrm{p}}}_{\mathrm{n}}\ast {\mathrm{r}}_{\mathrm{n}}\left)\right]}{\mathrm{F}{\mathrm{t}\mathrm{o}\mathrm{p}}_{\mathrm{n}}}}\hfill \\ +{\displaystyle \frac{{\left[\right({\mathrm{F}\mathrm{S}\mathrm{I}}_{\mathrm{F}\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{m}}}_1\ast {\mathrm{r}}_1)+\cdots ({{\mathrm{f}\mathrm{S}\mathrm{I}}_{\mathrm{F}\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{m}}}_{\mathrm{n}}\ast {\mathrm{r}}_{\mathrm{n}}\left)\right]}{\mathrm{F}{\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{m}}_{\mathrm{n}}}}+\left({{\mathrm{F}\mathrm{S}\mathrm{I}}_{\mathrm{F}\mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}}}_1\ast {\mathrm{r}}_1\right)\hfill \\ +\cdots \left({{\mathrm{F}\mathrm{S}\mathrm{I}}_{\mathrm{F}\mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}}}_{\mathrm{n}}\ast {\mathrm{r}}_{\mathrm{n}}\right)\hfill \end{array}$$

(9)

where FSI is the Flood Susceptibility Index; r is the Pearson correlation between Flood incidences and significant flood factors considered as weights (Table 3). FSIF_top: Topographic features. FSIF_clim: Long-term daily Antecedent, Precipitation Index (API), Long-term annual precipitation, Long-term daily and annual mean surface temperature; FSIF_other: Inverted Distance from Faults and earthquake hotspots (1950–2022), Inverted distance from road, Land cover factor based on the number of flood incidences per land cover type, Lithology factor based on the number of flood incidences per lithology type, Soil texture factor based on the number of flood incidences per soil texture type, Topographic wetness index (TWI) or compound topographic index (CTI); F_topn: Number (6) of topographic feature; F_climn: Number (3) of climate factors (the results of this equation is a number ranging from 0 to 1, which express the absence (0) and presence (1) probability of flood, FSI is the flood susceptibility index (FSI) with a probability value ranging from 0 (flood) to 1 non flood). In addition, the current and future flood Susceptibility Index were classified into five classes: No Flood (FSI < 0.60), Low (0.60 < FSI ≤ 0.65), Moderate (0.65 < FSI ≤ 0.657), High (0.7 < FSI ≤ 0.87), and Very High (FSI > 0.87) based on validation with the satellite-based flood map of May 03, 2023, using Sentinel-1 data.

FSI layer was normalized by using Equation (10) to generate a well-scaled FSI layer with values ranging from 0 to 1.

$${\mathrm{F}\mathrm{S}\mathrm{I}}_n={\displaystyle \frac{\mathrm{F}\mathrm{S}\mathrm{I}-{\mathrm{F}\mathrm{S}\mathrm{I}}_{min}}{{\mathrm{F}\mathrm{S}\mathrm{I}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{F}\mathrm{S}\mathrm{I}}_{min}}}$$

(10)

where FSI_n is a normalized FSI with values ranging from 0 to 1, which expresses the absence of low (0) and presence of high (1) probability of flood risk, respectively; FSI_max is the FSI maximum value; FSI_min is the FSI minimum value.

The FSI_n layer was classified into 5 flood classes below for easy identification of the flood hotspot areas that require much attention: Very Low FSI = 0–0.55, Low FSI = 0.55–0.6, Moderate FSI = 0.6–0.7, High FSI = 0.7–0.85, and Very High FSI = 0.85–1. The fitted logistic regression model was applied to the entire study area by passing raster values (predictor variables) into the model using a pixel-wise operation. The resulting output was a probability map (ranging from 0 to 1. Therefore, the classified map was exported in GeoTIFF format and then classified using ArcGIS Pro 3.5 software.

4. Model Validation

Ground truth points of the flood were identified from Sentinel-1 Synthetic Aperture Radar (SAR) imagery of 10 m resolution for the peak flood period from 2014 to 2024 across the country. The total collected points of flood and non-flood were 57,459 as shown in Figure 6 therefore 17,238 (30%) of all sampling data were used to validate logistic regression modeling.

The Receiver Operating Characteristic (ROC) curve (Figure 7) and the Area under the ROC Curve (AUC) validation results indicate that the present developed Python-based Logistic Regression Model categorizes areas susceptible to flood and stable areas with an AUC of 0.976, as illustrated in

Table 4. Pearson correlation of Flood Susceptibility Index in Rwanda based on Logistic Regression Model.

Area Under the Curve
Table Result Variable(s): FSI
Area	Std. Error	Asymptotic Sig.	Asymptotic 95% Confidence Interval
			Lower Bound	Upper Bound
0.976	0.001	0.000	0.974	0.978

.

The validation assessment based on Pearson correlation also indicated that this flood incidence-driven logistic regression model predicted well the FSI of Rwanda with a high r value of 0.0.976 (Table 4) and a p-value of 0.000.

As shown in Figure 7, the blue ROC curve represents the actual performance of the model, while the green diagonal line indicates the performance of a random classifier.

5. Results and Discussion

5.1. Flood Susceptibility Map of Rwanda

Figure 8 illustrates the spatial distribution of the flood susceptibility index across Rwanda, generated using the logical regression model that was performed in a Jupyter notebook using Python language for the baseline period of 1981 to 2017. The FSI was developed using a comprehensive multi-criteria analysis approach that incorporated twenty-two influential factors, including both hydrological and geo-environmental parameters.

One of the dominant drivers of flood susceptibility identified in this model is precipitation. The results clearly show a strong correlation between areas of high rainfall intensity and zones with elevated flood susceptibility scores. As visualized in the map (Figure 8), regions experiencing higher precipitation tend to have significantly higher FSI values. This trend is especially pronounced in the Western and Northern Provinces of Rwanda, where orographic rainfall is common due to the hilly terrain. Conversely, regions with relatively low rainfall, such as parts of the Eastern Province, exhibit lower susceptibility indices. This emphasizes the critical role that precipitation plays in surface runoff generation and potential inundation.

Slope is another important factor influencing the spatial distribution of flood-prone areas. Areas characterized by low slopes, particularly where the slope gradient is less than 5%, are found to be significantly more vulnerable to flooding. This is because flatter terrains hinder rapid water drainage and encourage the accumulation and ponding of runoff water, which leads to surface flooding. The highest susceptibility zones are, in fact, those where the slope percent approaches zero, such as wetlands, floodplains, and valley bottoms. This relationship between slope and flood risk is consistent with hydrological theory and observed patterns of flood damage in Rwanda. High slope areas, on the other hand, tend to facilitate rapid runoff and generally exhibit lower susceptibility scores unless they are accompanied by fragile soils and limited vegetative cover.

Land use and land cover (LULC) also emerged as critical parameters in flood susceptibility modeling. Among the different land cover types, wetlands were found to be the most vulnerable to flooding, contributing 75.3% of the highest susceptibility scores. Wetlands act as natural buffers that absorb excess water during rainfall events, but they are also the first to overflow when inundated. Moreover, wetlands in Rwanda are often under pressure from anthropogenic activities such as agriculture, settlement expansion, and infrastructure development. These activities reduce the water retention capacity of wetlands and increase the risk of flooding. In contrast, forested areas tend to have lower flood susceptibility due to better canopy interception, root absorption, and soil stabilization functions.

Another critical determinant of flood risk highlighted in the FSI model is proximity to rivers. The analysis shows that areas located closer to rivers exhibit significantly higher flood susceptibility scores. This is especially true in regions where rivers traverse low-lying areas, valleys, or floodplains. Water from higher elevations flows downstream toward these river valleys, and during periods of intense rainfall, the rivers often exceed their capacity, resulting in overbank flooding. Major rivers such as Nyabarongo, Nyabugogo, and Sebeya have been repeatedly identified as hotspots for flooding, particularly during the long rainy seasons. This observation is supported by reports from the Ministry in Charge of Emergency Management (MINEMA) and previous research. For instance MINEMA (2021) [74] documented recurring flood events along these river systems, often accompanied by loss of life and property. Studies by Benjamin (2015) [75] and Bizimana and Schilling (2009) [76] also corroborate these findings, pointing to the need for integrated river basin management to mitigate flood risks in downstream communities.

The FSI model further reveals that areas with zero percent slope and those situated within floodplains are at the highest risk of flooding. These areas are typically found in regions such as Bugesera, Rusizi, and the Nyabugogo floodplain, where topography is flat and drainage is poor. The high-risk zone is defined by slopes of less than 5%, while moderate-risk areas fall within the 5–15% slope category. This classification helps prioritize mitigation measures and resource allocation by national and local authorities. From a spatial planning perspective, the FSI map serves as a crucial decision-making tool.

It allows policymakers and emergency managers to identify zones where preventive measures such as improved drainage infrastructure, flood-resilient housing, and land use regulation should be implemented. The map also offers valuable insights for future urban development, especially in rapidly growing cities like Kigali, Rubavu, and Musanze, where urban sprawl is encroaching into flood-prone zones.

Figure 9 illustrates the spatial distribution of the Flood Susceptibility Index (FSI) across the surface of Rwanda during the baseline assessment period. The results reveal a dominant pattern where most of the country exhibits limited susceptibility to flooding, consistent with Rwanda’s diverse and predominantly rugged topography. According to the classification outputs, approximately 56.60% of the national territory falls within the very low susceptibility class, while 32.12% lies within the low susceptibility class. Together, these categories indicate that nearly 89% of the country is characterized by terrain that minimally supports flood occurrence.

This distribution is strongly influenced by Rwanda’s geomorphic characteristics. As shown in Figure 1, Rwanda is largely composed of mountainous and hilly landscapes, with steep slopes dominating much of the western and northern regions. These areas are less prone to flooding due to rapid surface runoff, limited water accumulation zones, and well-drained slopes. Flooding is therefore more likely to occur in areas where topographic conditions allow water to accumulate, such as flat terrain, valleys, and wetland systems. This explains why the majority of high-risk flood zones appear in lower-lying landscapes rather than in upland areas. Despite the predominance of low-susceptibility zones, the analysis also highlights the presence of critical flood-prone environments. Approximately 1.76% of the country’s surface is classified as high flood susceptibility, while 1.74% falls into the very high susceptibility class. Although these proportions are relatively small, they represent areas of significant environmental and socio-economic concern.

These highly susceptible zones are primarily concentrated in wetlands, floodplains, and lower slope gradient’s locations where hydrological conditions favor water retention and overflow during peak rainfall periods. One of the most important examples is the Nyabarongo wetland system, identified as the largest and most hydrologically active wetland in Rwanda. This wetland forms a major floodplain of low elevation, acting as a natural drainage basin for several major river systems. The rivers Nyabugogo, Mbirurume, Rukarara, and Mwogo converge toward or influence this basin, contributing to substantial surface water accumulation, especially during intense rainfall seasons. According to REMA (2014) [77], the Nyabarongo wetland plays a critical role in Rwanda’s hydrology but is also one of the areas most exposed to recurrent flooding because of its low-lying position and the extent of its drainage network.

These findings confirm that flood susceptibility in Rwanda is closely controlled by topographic and hydrological factors. Mountainous and steep regions generally remain safe from flooding, while wetlands, valleys, and gentle slopes form the key exposure zones. The results further underscore the need for targeted flood management interventions, particularly in the identified high-risk areas. Strengthening wetland protection, improving urban drainage, especially in zones influenced by the Nyabugogo river system and monitoring land-use changes in floodplains are essential measures for reducing long-term flood vulnerability. Overall, the FSI map presented in Figure 8 provides a comprehensive understanding of spatial flood patterns in Rwanda, offering a useful baseline for planners, policymakers, and disaster-risk managers. By clearly identifying both secure and highly exposed zones, the analysis supports evidence-based decision-making aimed at enhancing resilience and minimizing the impacts of future flood events.

5.2. Predicted Flood Susceptibility Map of Rwanda in 2030

Figure 10 presents the projected Flood Susceptibility Index (FSI) for the year 2030, derived from the annual precipitation forecasts of the Global General Circulation Model IPSL-CM5A-Medium Resolution (MR). This model is widely used in climate-impact assessments due to its robust representation of rainfall patterns and hydrological processes over East Africa. By integrating the model’s anticipated precipitation changes into the flood susceptibility framework, the spatial distribution of flood-prone areas reveals significant shifts compared to the baseline FSI shown in Figure 7. In contrast to current flood susceptibility conditions, the 2030 projections suggest that a considerable portion of Rwanda will experience an increase in susceptibility to flooding. The most notable changes appear in the Northern Province and specific high-risk environments of the Western Province, particularly around the Bugarama wetland. These shifts correspond with the model’s prediction of increased precipitation intensity and frequency in these regions by 2030. The North-West and South-West parts of the country are expected to receive higher rainfall compared to historical averages, thereby elevating the likelihood of surface runoff, riverine overflow, and localized flooding.

The North-Western region already characterized by steep topography, high elevation, and numerous streams, becomes especially vulnerable when overlaid with increased precipitation. The enhanced rainfall projected in 2030 is likely to generate rapid runoff into valley bottoms, intensifying the probability of flash floods and erosion-related flooding. Similarly, the South-Western areas, including wetlands adjoining the Bugarama lowlands, show a marked rise in susceptibility. The Bugarama wetland, located in one of Rwanda’s lowest elevation zones, becomes particularly exposed because it serves as a natural water-collecting basin. The model indicates that with increased precipitation expected in this region, the wetland will face higher inundation pressure, resulting in both a higher frequency and greater spatial extent of flood events.

Interestingly, while several regions show elevated susceptibility, the model also reveals areas where susceptibility decreases by 2030. The most significant reduction is observed in the Nyabarongo wetland system, especially in its central and downstream sections. This change is partly attributed to the projected decline in precipitation across parts of the Eastern Province, which impacts the overall hydrological inflow into the Nyabarongo basin. As the Eastern Province is projected to receive below-average rainfall, surface runoff and upstream contributions that feed the Nyabarongo wetland will be reduced. Consequently, the flood susceptibility in this wetland decreases relative to the baseline period.

In regions where rainfall is projected to increase, the model shows a corresponding amplification of flood susceptibility. Conversely, areas expecting reduced precipitation demonstrate a decline in susceptibility. These dynamics highlight the direct influence of climate-induced hydrological changes on future flood hazards in Rwanda. Overall, the 2030 FSI map provides critical insight into the evolving flood risk landscape under changing climatic conditions. The contrast between increased susceptibility in the North-West and South-West and reduced susceptibility in the Nyabarongo wetland and Eastern Province highlights the need for region-specific adaptation strategies. Decision-makers, urban planners, and disaster-risk managers should prioritize interventions in regions projected to experience heightened flood exposure. These may include strengthening early warning systems, implementing watershed management programs, reinforcing drainage infrastructure, and enhancing wetland conservation, especially in the Bugarama and volcanic highland catchments.

The findings presented in Figure 7 highlight the importance of integrating climate projections into national flood-risk planning. By anticipating where flood hazards are likely to intensify or diminish, Rwanda can better allocate resources and design adaptive strategies that safeguard vulnerable communities, ecosystems, and infrastructure by 2030 and beyond.

6. Conclusions

Flood susceptibility mapping is essential to delineate flood-prone areas and assess mitigation measures.

Rwanda is mostly dominated by ridges and plateaus including the Congo Nile with a topographic feature that is entirely hilly. The rainwater originating from the ridges flows towards the valleys which cannot effectively absorb and accommodate all the water owing to the increase in solid wastes from anthropogenic activities that clog culverts, and water drainage systems within the area. It showed that the Northern province is the part of the country with low flood susceptibility, which can be justified by the topographic nature and the geomorphological aspect of the country. The areas with an abundance of rainfall fell into highly susceptible classes. After modeling the Flood Susceptibility Index, we perceived that the areas with low distances to rivers, low elevation, and abundance of precipitation revealed a big influence on flood occurrences.

The government of Rwanda has launched the Ministry in Charge of Emergency Management (MINEMA), a practice of monitoring and managing disasters in Rwanda. Nonetheless, for the policy to be productive and sustainable, there is a great need to manage the increase in floods. Therefore, the following are suggested for the flood susceptible management in Rwanda:

• Rwanda’s growing population is expected to intensify pressure on land resources, potentially accelerating land degradation. It is therefore recommended that MINEMA collaborate closely with key institutions such as the National Land Authority (NLA), Rwanda Water Resources Board, the city of Kigali, and the Rwanda Housing Authority to develop a national urban master plan that fully integrates disaster risk considerations. Furthermore, strict regulations and penalties should be enforced for individuals who settle in high-risk flood zones, particularly in wetlands and riverine areas, in line with national urban planning guidelines.
• Rwanda receives abundant rainfall, particularly during the rainy seasons. Maximizing rainwater harvesting would help reduce pressure on underground water resources while also limiting the transport of sediments from high-elevation and steep-sloped areas to low-lying zones. This practice would contribute to better watershed management and reduce environmental degradation.
• It is recommended that Rwanda strengthen environmental education from primary school through undergraduate studies. Therefore, building ecological and hazard-awareness skills at an early age can improve community understanding of disaster risks.
• Integrate AI-driven flood susceptibility mapping into climate policy and adaptation planning: The government should use evidence-based AI approaches to identify high-risk flood zones, prioritize interventions, and strengthen national climate resilience frameworks. This integration will also help align local adaptation measures with international commitments, including the Paris Agreement and the Sustainable Development Goals.
• Strengthen disaster risk reduction and community preparedness: The government should utilize precise identification of flood-prone areas to implement proactive strategies, including early-warning systems, resilient infrastructure planning, and effective land-use regulation. These measures will reduce vulnerability and enhance preparedness, especially in rapidly urbanizing and environmentally sensitive regions.