Quantitative Assessment of Future Environmental Changes in Hydrological Risk Components: Integration of Remote Sensing, Machine Learning, and Hydraulic Modeling

Abstract

Floods are one of the most devastating natural hazards that have intensified due to land use land cover (LULC) changes in recent years. Flood risk assessment is a crucial task for disaster management in flood-prone areas. In this study, we proposed a flood risk assessment framework that combines flood vulnerability, hazard, and damages under long-term LULC changes in the Tajan watershed, northern Iran. The research analyzed historical land use change trends and predicted changes up to 2040 by employing a Geographic Information System (GIS), remote sensing, and land change modeling. The flood vulnerability map was generated using the Random Forest model, incorporating historical data from 332 flooded locations and 12 geophysical and anthropogenic flood factors under LULC change scenarios. The potential flood damage costs in residential and agricultural areas, considering long-term LULC changes, were calculated using the HEC-RAS hydraulic model and a global damage function. The results revealed that unplanned urban growth, agricultural expansion, and deforestation near the river downstream amplify flood risk in 2040. High and very high flood vulnerability areas would increase by 43% in 2040 due to human activities and LULC changes. Estimated annual flood damage for agriculture and built-up areas was projected to surge from USD 162 million to USD 376 million and USD 91 million to USD 220 million, respectively, considering 2021 and 2040 land use change scenarios in the flood-prone region. This research highlights the importance of land use planning in mitigating flood-associated risks, both in the studied area and other flood-prone regions.

1. Introduction

Floods are the most extensive global natural phenomena that cause noticeable damage to economic activity, people’s lives, the natural environment, property, and substructures [1]. Land use land cover (LULC) change, encompassing urban sprawl, deforestation, and agricultural land expansion, has substantial impacts on the hydrologic cycle, hydraulic flow, and climate systems [2,3]. These impacts contribute to increased flood frequency and intensity and consequently raise flood risk levels [4]. Accurately assessing the connection between LULC dynamic changes and flood risk provides an effective decision-making tool for risk mitigation [5].

Generally, flood risk encompasses three main components: hazards (flood extent and depth), vulnerabilities (probability of exposure), and flood consequences (loss of human life, property damage, and infrastructure damage) [6,7]. Prior research has addressed flood vulnerability, hazard, and damage as distinct components using a variety of methods [8,9]. These methods include physical modeling, probabilistic analysis, geospatial analysis, and multiple-criteria decision making (MCDM) tools, coupled with hydrological and numerical hydraulic models [10]. For instance, HA-Mim et al., in 2022, generated hazard exposure and vulnerability maps using the MCDM method and geospatial techniques in the Barguna district of Bangladesh [11]. Adnan et al., in 2020, assessed flood damages through the analysis of diverse flood recurrence intervals and LULC change scenarios; an increase in flood damages corresponding to shifts in both recurrence intervals and LULC scenarios was revealed [10]. Guido et al., in 2023, analyzed the potential impacts of flood events and identified vulnerable areas by considering the influence of LULC changes through hydraulic modeling (HEC-RAS) and geospatial analysis in a small watershed in the southeast United States [12]. Previous studies have made valuable contributions to assessing hazards, vulnerabilities, and flood consequences under LULC changes [13]. In fact, flood consequences arise from the interplay of hazards and vulnerabilities. Flood hazard, influenced by variables such as flood vulnerability, occurrence probability, velocity, depth, and duration, determine the extent of consequences [14]. Flood vulnerability directly influences an area or population’s risk level during a flood hazard event, intensifying potential impacts and emphasizing consequences like damage and loss [13]. However, no research has developed a comprehensive framework that effectively and simultaneously integrates the identification of flood vulnerability, hazards, and damages under long-term LULC changes [5].

Furthermore, flood risk assessment may have limitations in providing reliable and accountable outcomes for flood control when dealing with large and complex datasets [15]. In this regard, the machine learning (ML) models and data mining technique models have gained attention for their ability to excel in creating vulnerability maps by efficiently handling complex relationships, processing large datasets, and providing improved accuracy, faster computational times, and reduced model development costs [10]. Notably, significant ML models, such as Random Forest (RF) [16], Linear Regression [17], Support Vector Machine (SVM) [18], Neural Networks (NNs) [19], Frequency Ratio (FR) [20], and the Bayesian Generalized Linear Model [21], have been utilized to forecast various natural hazards. Hence, it is essential to employ efficient ML models and data mining techniques to assist urban planners and decision-makers in identifying high-risk, flood-prone areas under complex data.

Therefore, this study aims to develop a comprehensive framework that integrates multiple techniques and models, including ML, hydraulic modeling, remote sensing (RS), GIS, and Land Change Modeler (LCM), to assess the impacts of long-term LULC changes on flood vulnerability, hazard, and consequences in the Tajan watershed of northern Iran. This framework uniquely combines flood vulnerability, hazard, and damage assessments under long-term LULC change scenarios into a single modeling approach. By utilizing advanced technologies that reduce uncertainties, it provides a comprehensive and dynamic assessment of flood risk, surpassing previous studies that have examined these aspects separately. Using RS-GIS and LCM, we predict future LULC changes in 2040 based on historical trends. The widely used ML model, Random Forest, and the HEC-RAS hydraulic–hydrology model are utilized to map flood vulnerability and hazards, considering the long-term dynamic changes in LULC. Global depth–damage curves and functions under changing LULC conditions are used to calculate the negative consequences of flooding, including direct financial damage to flooded agricultural land and built-up areas. This study will (i) identify high-risk areas, flood extent, and depth under different LULC change scenarios and varying time intervals, and (ii) estimate the potential flood damage costs in residential and agricultural areas, considering long-term LULC changes within flood-prone regions and over different time intervals. The findings will aid urban planners and policymakers in effectively managing and mitigating the consequences of flooding, particularly in rapidly developing countries dealing with urban growth and flood damage.

2. Cases Study

The Tajan River watershed is sited between 53°05′ and 53°18′ Eastern longitude and 36°09′ and 36°29′ Northern latitude in Mazandaran Province, Iran. As shown in Figure 1, it passes through 77 villages, covering an area of roughly 3810 km² with a length of 140 km. The climate of the area is extremely humid and generally mild, with an annual temperature of nearly 14 °C and an average yearly rainfall precipitation of 830 mm; maximum and minimum precipitation occur in autumn/winter and summer, respectively. About 65–70% of annual rainfall takes place during the September–December period [22]. The altitude in the region varies from −2 m at the outlet to 3720 m above mean sea level. This area’s dominant land cover types include rangeland, forest, agricultural, waterbodies, and built-up areas. The economy of the study area relies mainly on rice cultivation and its associated supply chains [23]. The Tajan River regularly experiences severe floods with high discharge, resulting in damage to built-up areas and agricultural lands every year [22]. For example, the 2019 flood in Mazandaran Province resulted in damages amounting to USD 4.7 billion across agricultural, industry, residential, commercial, and road properties, making it as the most costly natural disaster in Iran [24]. Despite the severity of such floods, research on flood risk management in this watershed remains limited [19,25].

3. Methodology and Data

This research was conducted in the following four main stages: (i) LULC map development and future prediction; (ii) flood vulnerability assessment based on LULC change scenarios; (iii) flood hazard assessment for different return periods under LULC change scenarios; (iv) flood consequences assessment in terms of economic damages. Figure 2 provides a visual representation of the methodology framework.

3.1. LULC Maps Development and Future Prediction

3.1.1. Processing Remote Sensing Images to Generate a LULC Map

Spatial–temporal changes in different land cover categories are revealed using scenes from the Landsat 5, Landsat 7, and Landsat 8 satellites [2]. Various satellite images, including those from sensors such as ETM+, TM, and OLI, with a resolution of 30 m, were obtained from the United States Geological Survey (USGS) Earth Explorer. These images were utilized to create LULC maps for three distinct years: 2001, 2011, and 2021. Detailed information about the features of the different Landsat sensors can be found in

Table 1. Satellite image features.

Landsat Sensor	Image Generated Date	WRS Row	WRS Path	Spatial Resolution	Number of Bands Used	Datum
ETM+	30 July 2001	35	163	30	6	WGS84
TM	28 March 2011	35	163	30	6	WGS84
OLI	27 June 2021	35	163	30	8	WGS84

.

In this study, we employed the maximum likelihood algorithm to analyze the LULC categories in the Tajan watershed. To create the LULC maps, we utilized 6 to 8 bands from Landsat images for each year. Radiometric and topographic normalization, as well as geometric correction, were applied using ArcGIS tools to ensure accurate LULC classification and to minimize the impacts of factors such as atmospheric conditions, solar angle, and sensor view angle [26]. Following the pre-processing steps, the maximum likelihood classification method was employed, utilizing nearest neighbor analyses with 150 training samples, to categorize the LULC maps into five classes of land use and land cover, as presented in Equation (1). Accuracy assessment of the classification results is crucial in aerial imagery analysis. We used two uncertainty matrices, the overall accuracy test and the kappa coefficient matrix, to estimate the accuracy of the generated maps [27].

$$P(clas{s}_i|data)={\displaystyle \frac{P\left(data|clas{s}_i\right)·P\left(clas{s}_i\right)}{P\left(data\right)}}$$

(1)

where $P(clas{s}_i|data)$ is the probability of class i given the data, $P\left(data|clas{s}_i\right)$ is the likelihood of the data given class i, $P\left(clas{s}_i\right)$ is the prior probability of class i, and $P\left(data\right)$ is the probability of the data.

3.1.2. Prediction of LULC Changes and Model Validation

The Land Change Modeler (LCM) within the TerrSet geospatial monitoring software 2020 was utilized to measure the biodiversity effects in order to investigate and predict land cover changes. This model incorporates several techniques, including Markov chain matrices, artificial neural networks (ANNs), logistic regression (LR), and multilayer perceptron (MLP) [28]. To generate future LULC changes using LCM, we followed four steps: 1. Calculation of the amount and spatial trend of changes. 2. Estimation of transition potential and identification of essential factors driving LULC changes. 3. Modeling of LULC changes. 4. Model validation and accuracy assessment. The validation and agreement were assessed using the overall accuracy (OA) and kappa index between predicted and actual maps. Detailed explanations for these steps are provided in Appendix A.

The overall accuracy (OA), defined in Equation (2), was calculated by dividing the number of correctly identified pixels by the total number of categorized pixels in the error matrix [29].

$$A={\displaystyle \frac{1}{N}}\sum P_{ii}$$

(2)

where N represents the total number of accuracy pixels, and $\sum P_{ii}$ denotes the summation of the major diagonal of elements in the error matrix. The kappa index measures the degree of agreement between the predicted and actual maps when considering all elements of the error matrices. It is calculated using Equation (3):

$$Kappa={\displaystyle \frac{P_i-P_c}{1-P_c}}\times 100$$

(3)

where $P_i$ represents the observed agreement in the error matrices (overall accuracy), while P_c represents the expected agreement. The kappa index could range from −1 to +1, with a value of 1 indicating perfect agreement. A value of 0 indicates random classification, while a negative value implies agreement worse than random [30].

3.2. Flood Vulnerability Assessment Method

To create flood vulnerability maps, we utilized the Random Forest model and a GIS application, incorporating flood inventory data and flood influential factors. The Mean Decrease Accuracy (MDA) method, known for estimating the weights of flood-influencing factors, was employed to select the relevant input variables for modeling [31]. The process involved the spatial modeling of geo-environmental factors to determine flood vulnerability and locations, using testing and training datasets. Subsequently, the vulnerability maps were generated using the RF model and the training datasets. To assess the impact of land use and land cover (LULC) changes on flood probability, the RF model was trained separately for two LULC scenarios: 2021 and 2040. To validate the RF models, different statistical measurements were applied to the testing and training datasets. A comprehensive explanation of the methods and data used in generating flood vulnerability maps based on the two LULC change scenarios is provided in the subsequent section.

3.2.1. Inventory and Mapping of Flood

A flood inventory map is a crucial tool that offers valuable insights into areas that have been affected by flooding in different historical periods. It serves as a valuable resource for researchers for identifying flood patterns and predicting future occurrences [23]. The flood distribution map of the Tajan watershed, shown in Figure 1, was obtained from the Management of Land and Water Resources Company of Iran. The flood inventory includes 332 recorded flood events that occurred between January 1990 and July 2021. These events were compiled using various sources, such as satellite imagery from Google Earth and field surveys. Considering past flood patterns, current flood risk serves as a valuable predictor for future occurrences [32]. To facilitate modeling and analysis, 332 non-flooded locations were randomly selected in ArcGIS 10.5, creating a dichotomous dependent variable for the machine learning models. The dataset was divided into training (70%) and testing (30%) sets using random selection in Python with the Scikit-learn library, following the common practice in machine learning to allocate sufficient data for training and evaluate model performance [33]. Notably, the selection of non-flooded points deliberately focused on areas with lower flood risk, including hills, forested lands with gentle slopes, and mountains. This intentional choice ensured unbiased representation and enhanced the performance of machine learning models [34]. By evaluating the models on the testing set, their performance could be assessed, and the sensitivity of the results to the data split could be determined. This approach offered insights into the models’ generalization capabilities and their ability to predict flood susceptibility beyond the training data [35].

3.2.2. Preparing Flood Influencing Factors

Collecting and constructing influential data on floods is an essential part of simulations. Considering the results published in previous research, we selected twelve influencing factors to predict future floods based on the study area’s topography, geology conditions, hydrology, and climatic and geo-environment conditions [36,37,38]. Figure 3 depicts the selected factors utilized in our analysis, incorporating remote sensing techniques. Spatial data representing these factors were converted into a 30 m resolution raster format using ArcGIS and nearest neighbor interpolation. The natural break method was employed to classify the data into distinct categories, facilitating their seamless integration into the flood susceptibility modeling framework [33]. The topography factors included slope (SL), aspect (AS), altitude (AL), the Topographic Position Index (TPI), and the Terrain Ruggedness Index (TRI). The geology factors consisted of the information value of soil texture (ST) and lithology (LI). Drainage density (DD), distance from river (DR), rainfall (RA), and the Topographic Wetness Index (TWI) made up the hydrology factors, and the geo-environment factor was the land use land cover (LULC). Furthermore, we employed the multicollinearity (MC) test to evaluate the linear relationships among various influencing factors in the context of flood vulnerability assessment. Comprehensive explanations of the flood-influencing factors and the MC test can be found in Appendix B.

3.2.3. Random Forest (RF) and Validation

RF is a supervised machine learning algorithm that combines multiple decision trees to achieve a single result, making it suitable for regression and classification tasks [39]. Before training the model, three main parameters need to be adjusted: node size, the number of trees, and the number of sampled features (variables). The number of features can range from one to the entire set of variables, while the number of trees is typically set between 500 and 1000 [40]. Each decision tree in the Random Forest is trained on a bootstrap sample from the training set, with one-third (30%) of the data reserved as an “out of bag” (OOB) sample for testing. Feature bagging is then applied to introduce additional randomness, increasing dataset diversity and reducing correlation between decision trees [41]. To assess the importance of variables in predicting flood risk, Random Forest measures their impact. In this study, the importance ranking of flood influencing factors was evaluated using the MDA method. MDA calculates the accuracy reduction associated with the presence or absence of specific variables, and the importance of independent factors is evaluated through permutation [42]. The importance factors were evaluated using the MDA method, as presented in Equation (4) [43]:

$$\mathit{Variable}\mathit{Importance}:VI\left(x_i\right)=\frac{1}{n_{tree}}{\sum }_{t=1}^{n_{tree}}{\displaystyle \frac{{\sum }_{iϵOO}I\left(y_i=f\left(x_i\right)\right)-{\sum }_{iϵOOB}I\left(y_i=f\left(x_i^j\right)\right)}{\mid OOB\mid }}$$

(4)

The OOB (out-of-bag) method is used in Random Forest to estimate the prediction error of each training sample. It is a subsampling technique that employs bootstrap aggregation (bagging) by creating training samples through subsampling with replacement. OOB accuracy is measured before and after the permuting variable x_j, and the difference is calculated to obtain a more accurate prediction value [31]. For each tree t, where t ∈ {1, 2, 3…, n_tree}, the variable importance of x_j is determined by averaging the difference between the predicted class before permuting x_j (y_i = f(x_i)) and after permuting variable x_j (yi = f(x_ij)) for a specific observation i [42]. Indeed, this study assessed the performance of the Random Forest model by using various statistical measures to evaluate its ability to generate precise flood vulnerability maps. These measures encompass the Sensitivity Test (SST), Specificity (SPF), Positive and Negative Predictive Values (PPV and NPV), Root-Mean-Squared Deviation (RMSD), Coefficient of Determination (R²), and Receiver Operation Characteristics (ROC-AUCS). Detailed descriptions of the validation methods and the statistical measures, including their equations, can be found in Appendix C.

3.3. Flood Hazard Assessment Methods

The evaluation of flood hazards involves conducting hydrological simulations of floods across multiple return periods. For this study, the Hydrological Engineering Center-River Analysis System (HEC-RAS v6) tool was utilized to calculate flood depth and velocity in coastal and fluvial regions [14,44]. The selection of the study area for the flood hazard assessment was based on several factors, including historically flooded locations, the increase in urbanization and cultivation along the river, and the observed rise in stream flow in recent years [45]. These factors collectively led to the identification of a specific part of the study area, covering approximately 205 km², as a highly flood-susceptible zone for conducting the hazard assessment (shown in Figure 4). It is important to note that the chosen area represents a subset of the overall study region. Due to limitations in data availability, it was not feasible to include the entire region in the analysis. However, the selected area was deemed most relevant for assessing flood hazards based on known flood occurrences and the influence of urbanization and agricultural activities along the river [23]. To initiate the assessment, maximum discharge data recorded by the Sari Water Resources Department of Iran over the past 30 years (1989–2020) were collected from hydrological stations located upstream (Soleiman-Tange) and downstream (Balakoola) of the river (refer to

Table 2. Detailed information of the hydrological stations.

Station Name	River	Station Code	Latitude	Longitude	Elevation (m)
Soleiman-Tange	Tajan	13019	36-15-12	53-13-11	400
Balakoola	Tajan	13035	36-27-31	53-07-15	115

and Figure 5). Flood frequency analysis (FFA) was then conducted using the Easyfit application, which utilizes probability distribution functions such as Log Pearson type 3 (LP3), to establish the relationship between the magnitude of extreme runoff or river flow events (Q) and their frequency of occurrence (also known as recurrence interval or return period) (T). The analysis considered various return periods, including Q2, Q5, Q10, Q25, Q50, Q100, Q500, and Q1000 years. Assessing a 1000-year flood event was particularly crucial for effective risk management, given the occurrence of excessive discharge in 2019 and the expansion of rural areas and agricultural lands along the river [14]. Subsequently, the HEC-RAS (1D/2D) combined model, a widely used software tool for flood depth and velocity calculations in coastal and fluvial regions, was employed to estimate the depth and intensity of flood inundation for each pixel (12.5*12.5) in the study region. The simulation process involved extracting river geometry data from the 12.5 m digital elevation model and creating bank lines, centerlines, flow path lines, and cross-section lines. Floodplains were then modeled for the main branch of the river, consisting of 78 cross-sections along a 40.63 km stretch of the Trajan River, with an average distance of approximately 0.59 km per section.

It should be noted that, in this study, manual calibration was required for the HEC-RAS model since it lacks an automatic calibration option. This calibration focused on Manning’s roughness coefficient (N value), a crucial parameter within the HEC-RAS model [46]. We adjusted Manning’s N values for both the main channel and side banks in the upstream and downstream regions. The calibrated values for the main channel were set at (0.04, 0.035, 0.05), while for the side banks, they were (0.045, 0.05, 0.06). Boundary conditions and flow types (sub-critical and super-critical) were determined based on the normal depth and slope profiles upstream and downstream. We conducted a steady-state analysis for all flood flow periods, involving a statistical assessment of 20 out of 30 recorded events from 1989 to 2020 to evaluate differences between simulated and observed water depths. This analysis employed metrics like the correlation coefficient (R), the Nash–Sutcliffe efficiency (NSE), and Root-Mean-Squared-Deviation (RMSD) [13,31].

3.4. Flood Consequences

To estimate the direct financial damages to agricultural land with crops and the built-up area based on land use and land cover (LULC) change scenarios, Equation (5) was employed in different cells and for the total cell area as follows:

$$D_j=\left(\sum _{i=1}^n{x}_i\times f\left(x_i\right)\right)\times A$$

(5)

where D represents the total damage sustained throughout flood recurrence interval j (in million USD),$x_i$ represents the depth of flood in pixel i (m), $f\left(x_i\right)$ represents the function of damage in pixel i, and A represents the pixel area [47]. Accordingly, the adopted depth–damage curve at a continent level presented by [48] has been applied to calculate the perceptible flood damages to agricultural land and the built-up area, as shown in Figure 6. The flood depth–damage curves demonstrate the correlation between flood depth and total or relative economic loss, which directly affect financial losses [5]. While several factors influence the financial impact of a flood, such as inundation time, velocity and wave impacts, contamination, water temperature, and debris, many studies have predominantly considered only inundation depth and extent as influencing factors [14].

In this study, the maximum damage ratios from the depth–damage equation were converted from Euros to US dollars at a conversion rate of 1 Euro = 1.13 US dollars.

Table 3. The cost of max-damage to the built-up and agriculture land with the crops in two LULC scenarios.

Scenario	LULC Type	Max Damage
LULC (2021)	Built-up area	60 USD/sq.m.
	Agricultural land and crops	863 USD/hectare
LULC (2040)	Built-up	160 USD/sq.m.
	Agricultural land and crops	2760 USD/hectare

shows the cost of maximum damage considered to the built-up and agricultural land with the crops based on LULC scenarios, which is implemented from earlier research [48]. In this research, the costs of the built-up area and agricultural land with dominant rice crops were taken into account as 60 USD/sq. m. and 863 USD/hectare, respectively. The annual GDP value growth rates in rural areas for built-up areas and agricultural land were assumed to be 4.5% and 5.5%, respectively [14]. Furthermore, the max-damage in the next 20 years (2040) was expected to be 160 USD/sq.m. and 2760 USD/hectare for the built-up area and agricultural land with crops, respectively.

To assess flood damage with a spatial resolution of 30 m, ArcGIS was used to measure damage for nine flood recurrence intervals (2, 5, 10, 20, 50, 100, 200, 500, and 1000 years) under two LULC scenarios: 2021 and 2040. In order to match flood depth with the corresponding LULC type in each pixel, the inundation maps (discussed in Section 3.3) were overlaid on the land use maps in ArcGIS. Finally, the total flood damage at the pixel level was estimated using Equation (5) in MATLAB 23.2 software.

Subsequently, the risk was assessed by evaluating the expected annual damage (EAD) in both LULC change scenarios. The EAD was estimated as the area under the curve, as presented in Equation (6) [10].

$$\mathrm{EAD}=\iint D\left(P\right)dPdA$$

(6)

In this equation, D(P) represents the damage that happens with a high possibility (P) each year, which is inversely related to the flood recurrence interval (T). The variable A represents the total area. Considering all return times between a high and low probability of flooding provides an accurate estimate of risk because the choice of return periods affects flood risk calculations [13]. To visualize the relationship between flood damage and coupled occurrence probabilities, an EAD curve was constructed. The occurrence probabilities of 0.001 were determined to be the upper and lower boundaries of the probability curve.

4. Results

4.1. Predicted LULC Map and the Change Assessment

This study utilized maps from 2001, 2011, and 2021 were used to project the LULC alterations up to 2040. The historical and recent maps of the Tajan watershed were categorized into five classes: built-up, agricultural, waterbody, forest, and rangeland (see Figure 7). The accuracy of the classified LULC images was evaluated using a Neural Network classification algorithm, and the results are presented in

Table 4. The accuracy of the land use land cover classification of images in different years.

Assessment Indexes	LULC2001	LULC2011	LULC 2021
Cohen’s kappa index	0.81	0.83	0.81
Overall accuracy (%)	88.74	92.67	90.81

, which shows the acceptable accuracy rates measured by the kappa coefficient and overall accuracy tests. The MLP model employed in the LULC conversion sub-model demonstrated satisfactory estimates across all sub-models, as indicated by the findings presented in

Table 5. General features of various LULC conversion models from 2011–2021.

Sub-Model	Independent Factors (Numbers)	Factor’s Momentum	Interactions	Number of Hidden Layers	Accuracy Rate (%)
Forest–Agriculture	5	0.35	10,000	7	97.3
Forest–Water Body	5	0.20	6000	7	81.7
Forest–Built-Up	5	0.27	8000	5	86.7
Agriculture–Built-Up	5	0.32	10,000	6	84.3
Rangeland–Agriculture	5	0.24	7000	6	96.5

. Notably, the sub-models for the conversion from forestland to agricultural land and from rangeland to built-up areas achieved the highest accuracy rates of approximately 97.3% and 96.5%, respectively, making them the most precise sub-models among the generated models. Additionally, the LULC change map for 2040, projected using the Markov chain and LCM models, is depicted in Figure 8. The results indicate a satisfactory level of accuracy in the model’s predictions of future LULC changes within the study area, achieving an overall accuracy of 0.86 and a kappa coefficient of 0.89.

Table 6. Alterations in LULC area at various periods from 2001 to 2040.

LULC Type	LULC Area till 2001	LULC Area till 2011	LULC Area till 2021	LULC Area till 2040	Rate of LULC Changes 2021–2040 (Km2)	Percentage of Changes 2021–2040 (%)
Built-up	30.32	39.64	48.5	64.34	15.84	+0.42
Agriculture	417.7	597.9	766.9	1040.2	273.3	+7.2
Forest	1990.1	1876.97	1788.6	1700.20	−88.4	−2.3
Water Body	3.6	3.9	4.1	4.6	0.5	+0.01
Rangeland	1360	1283.3	1193.6	992.37	−201.23	−5.88

presents the alterations in different LULC types during various periods calculated by the Markov chain matrix. The results reveal that forestland was the most extensive land cover, occupying about 50% of the entire area in 2001. However, it is projected to experience a significant decline in future changes. Comparing changes between the LULC maps for 2021 and 2040 shows that the forest area will shrink by 88.4 km² in the next 20 years. The most dynamic classes are agricultural lands and built-up areas, which will increase by approximately 273.3 km² and 15.84 km², respectively, by 2040. The results show that the built-up area and agricultural land are expected to double by 2040. The map illustrates that a majority of the expansion in built-up and agricultural land has occurred along the river, which are critical areas prone to flood exposure. Rangelands, including grasslands, shrublands, and woodlands, are projected to decrease significantly by 227.83 km² from 2021 to 2040. On the other hand, the water body experienced a slight increase of about 1.2% from 2001 to 2040 primarily due to changes in rainfall patterns [49].

4.2. Flood Vulnerability Assessment and Mapping

4.2.1. Flood Influencing Factors for Vulnerability Mapping

Before ranking all flood contributing factors in the watershed, a multicollinearity assessment was conducted using TOL (tolerance) and VIF (variance inflation factor) statistics to evaluate the presence and strength of correlation among the flood influencing factors in the watershed. Based on the results obtained from

Table 7. Multicollinearity assessment test.

Flood Influencing Factors	VIF	TOL
Slope	1.02	0.97
Aspect	1.4	0.94
Altitude	3.72	0.20
TPI	1.81	0.87
TRI	1.64	0.62
Information value of Soil Texture	2.16	0.46
lithology	2.07	0.51
Drainage Density	1.39	0.78
Distance from River	1.5	0.69
Rainfall	3.69	0.28
TWI	1.43	0.81
LULC	1.09	0.94

, the VIF values range from 1.02 to 3.72, and the TOL values range from 0.30 to 0.97. Typically, VIF values below 10 and TOL values above 0.1 are considered acceptable thresholds, indicating low multicollinearity. These values suggest that multicollinearity is not a significant concern among the flood-influencing factors in this study. The VIF and TOL values for all predictor variables fall within acceptable ranges, indicating low correlation between variables [25]. Therefore, multicollinearity is not a problem in this study, and the predictor variables can be considered independent and suitable for further analysis.

In this study, the MDA selection method was employed to identify and eliminate irrelevant factors, thereby reducing uncertainty in the prediction models using the RF machine learning algorithm. Figure 9a,b depict the ranking of flood influencing factors’ importance based on the RF machine learning model for the two LULC scenarios, 2021 and 2040. The findings indicate that the distance from the river (DR) emerges as the most crucial predictor of flood occurrence, exhibiting the highest values (W = 0.45, 0.47), while the soil texture (ST) factor demonstrates the lowest values (W = 0.03, 0.2) among all the input variables in both the 2021 and 2040 scenarios. In the 2021 LULC scenario, the most significant factors are distance from the river (DR) (W = 0.45), altitude (AL) (W = 0.31), land use land cover (LULC) (W = 0.23), slope (SL) (W = 0.19), drainage density (dd) (W = 0.17), and rainfall (RA) (W = 0.13). However, the rankings of important factors change in the 2040 LULC scenario. Distance from the river (DR) assumes the highest importance (W = 0.48), followed by land use land cover (LULC) (W = 0.39), altitude (AL) (W = 0.28), slope (SL) (W = 0.20), rainfall (RA) (W = 0.19), and drainage density (dd) (W = 0.15).

4.2.2. Validation of Flood Vulnerability Models

The flood vulnerability model was validated using the Receiver Operating Characteristic (ROC) curve and several statistical indices listed in

Table 8. The validation of flood vulnerability models through standard numerical indices.

Index	Training Dataset (Scenario 2021)	Validation Dataset (Scenario 2021)	Training Dataset (Scenario 2040)	Validation Dataset (Scenario 2040)
PPV (%)	0.82	0.79	0.87	0.83
NPV (%)	0.87	0.87	0.90	0.89
SST (%)	0.91	0.88	0.92	0.91
SPF (%)	0.85	0.78	0.89	0.87
R2 (%)	0.89	0.89	0.91	0.90
RMSD	0.33	0.37	0.28	0.31
AUC (%)	0.92	0.90	0.95	0.92

. These assessments evaluated the model’s ability to identify areas at high risk of flooding. The ROC curve visually displays the match between observed flood points and the model’s predictions, with the area under the curve (AUC) providing a quantitative measure of accuracy. Standard numerical indices were also used to assess the reliability of the model on both the training and testing datasets. Table 8 and Figure 10 present the model’s performance through parameters such as Positive Predictive Value (PPV), Negative Predictive Value (NPV), Sensitivity and Specificity Tradeoff (SST), Specificity (SPF), R², Root-Mean-Squared Deviation (RMSD), and AUC for the training and validation data. The RF model trained on the 2040 LULC scenario exhibited higher values for PPV (0.87%), NPV (0.90), SST (0.92), SPF (0.89), R² (0.91), and AUC (0.95%), indicating improved efficiency compared to the 2021 LULC scenario. Conversely, the RF model showed lower RMSD (0.28) for the 2040 scenario compared to 2021 (RMSD = 0.33). The validation dataset also yielded higher values for PPV (0.83), NPV (0.89), SST (0.91), SPF (0.87), R² (0.90), AUC (0.92), and lower RMSD (0.31) when considering the future scenario. In contrast, the validation data based on the 2021 scenario exhibited slightly lower values for PPV (0.79), NPV (0.87), SST (0.88), SPF (0.78), R² (0.89), AUC (0.90), and higher RMSD (0.37). These results confirm that the RF model accurately predicts the flood vulnerability pattern in the study area, both historically and under future conditions [16].

4.2.3. Flood Vulnerability Mapping

The flood vulnerability maps of the study area presented in Figure 11a,b were generated by the RF machine learning model based on the impacts of historical and future LULC change scenarios. The flood vulnerability map is classified into five categories: very low, moderate, high, and very high vulnerability, utilizing the natural break classification method in ArcGIS.

From Figure 12a,b, it can be detected that 13.8%, 27.5%, 22.7%, 22%, and 14% of the study area are placed in the very low, low, moderate, high, and very high flood vulnerable zones in the 2021 LULC scenario. In comparison, the future 2040 LULC scenario shows that 17.4%, 18.6%, 21.2%, 23.2%, and 19.6% of the study area are located in very low-, low-, moderate-, high-, and very high-risk vulnerable zones, respectively. These percentages represent the proportion of the total study area that lies within each flood vulnerability category. The maps for both scenarios reveal that high and very high flood vulnerability zones are predominantly located along the main Tajan River and its tributaries. Conversely, low and very low vulnerability zones are concentrated in the high-elevation areas of the watershed’s southwest. The moderate-vulnerability zones in both scenarios are mostly found downstream and in areas with gentler slopes.

4.3. Flood Hazard Mapping

The flood hazard mapping analysis provides crucial insights into the potential flood risks and associated damages within the study area. Using the HEC-RAS flood model, simulations of flood inundation, extents, and flow velocity were conducted for design events with return periods ranging from 2 years to 1000 years.

Figure 13a displays the flood inundation depth and extent map of the selected portion of the Tajan watershed (see Figure 4) for the Q1000 peak discharge at a 30 m topographic resolution. The flood inundation map reveals that the lowland areas adjacent to the Tajan River in the middle and downstream parts are characterized by very high flow depths. Furthermore, Figure 13b presents the percentage of built-up land, agricultural land, and forestland located in the selected portion of the Tajan watershed (Section 3.3) for the years 2021 and 2040. However, in this study, we are only looking at the impact of flooding on built-up areas and agricultural land with crops.

Figure 13c illustrates the relationship between maximum simulated peak discharges and flood return periods, along with the corresponding maximum water depth. The discharge rate is primarily influenced by the stream channel’s depth, width, and cross-sectional shape. As the return period increases, the discharge increases linearly, resulting in a larger area being exposed to flood risks and damages, consistent with previous research [50,51]. The maximum inundation depth varies from 2.2 to 4.6 m for a 2- to 10-year return period and ranges from 6.14 to 15.12 m for return periods of 25, 50, 100, 200, 500, and 1000 years. Figure 13d indicates that the flood inundation area increases linearly with higher return periods. The agricultural lands situated along the riverside, covering a significant portion of the watershed, are particularly vulnerable to flood risks, leading to substantial damage. On the other hand, the built-up areas, occupying a smaller portion of the watershed and located further away from the river, are more susceptible to flood intensity. It is worth mentioning that the model calibration involved adjusting Manning’s roughness coefficient to achieve the best agreement between observed and simulated inundation data. Figure 14 demonstrates the favorable agreement between observed and simulated water depths, with a Coefficient of Determination exceeding 0.90 at both stations for 20 reviewed events. The hydraulic models demonstrate strong performance, indicated by stage calibration statistics (see

Table 9. Model performance statistics during calibration and validation at two stations.

Performance	Stations	NSE	RMSD	R
Calibration	Soleiman-Tange	0.80	0.015	0.91
	Balakoola	0.79	0.018	0.90
Validation	Soleiman-Tange	0.83	0.019	0.89
	Balakoola	0.82	0.016	0.88

) with R values above 0.9, NSE values of at least 0.79, and RMSE below 0.015 at both upstream and downstream stations. Similarly, Table 9 presents the statistical indicators used to assess the models’ performance in flow validation at both stations.

4.4. Flood Damage Analysis

This study performed a flood damage analysis to evaluate the financial impact of flooding on affected areas, focusing on agricultural lands with crops and built-up regions. The analysis applied the flood depth–damage function (Equation (5)) and considered two LULC scenarios: 2021 and 2040.

To calculate the area affected by flooding, the flood hazard grid maps (depth and extents) for different return periods and the LULC maps of 2021 and 2040 were imported into the ARCGIS 10.5 software. By using geometric intersection tools, the affected land cover types in the study area were analyzed against the simulated flood extent for each return period. The total cost of flood damage was then determined by combining the depth of flooding in each cell, the affected area, and the cost of maximum damage (see Table 3) based on the LULC scenarios for 2021 and 2040.

Figure 15 presents the total calculated damages resulting from floods for various return periods under the two LULC scenarios. The analysis revealed that agricultural lands are more vulnerable to flooding due to their proximity to the river, and damage to these regions occurs across all flood return periods (Q2 to Q1000). In contrast, significant damages to built-up areas were observed only in return periods of 25 to 1000 years. The results show an increasing trend in damage costs due to flood depth and flow intensity, as influenced by variations in the return period and LULC scenarios.

The calculated average damage to built-up areas in all flood return periods (Q25 to Q1000) WAS USD 117.43 million in 2021 and is projected to rise to USD 287.24 million in 2040. Additionally, the estimated damages to agricultural lands in 2021 across all flood return periods (Q2 to Q1000) amounted to USD 162.82 million, which is expected to increase to USD 452.10 million by 2040. The most significant flood damage to agricultural lands is estimated to be USD 102.33 million, with a Q1000 return period in the 2040 scenario, attributed to a noticeable increase in cultivated lands. Similarly, the greatest flood damage to the built-up area is estimated to be USD 81.44 million with a Q1000 return period in the 2040 scenario. The observed increase in flood damage from 2021 to 2040 is attributed to the considerable alteration of LULC, resulting in an increase in agricultural land and built-up areas.

Furthermore, the probability of the exceedance curve was derived from the flood damages plot, considering the probabilities of exceedance, and the expected annual damage (EAD) was calculated from the area under the curve. Figure 16 illustrates the probability of the exceedance curve, and Figure 17 shows the calculated EAD, which signifies the influence of LULC alteration on flood risk. The results in Figure 17 indicate that flood damage varies significantly between the maximum and minimum exceedance probabilities in both scenarios. In 2021, when the probability of exceedance is less than 0.02, the maximum average damage to agricultural land increases to USD 34 million with a steeper slope. When the probability of flooding is more than 0.02, the damage decreases with a gentle slope to USD 8 million. Similarly, damages of USD 13 million were calculated for the built-up area in 2021 when the exceedance probability was more than 0.02, while the average annual maximum damage increased to USD 30 million when the exceedance probability was lower than 0.02.

In the 2040 scenario, damages to agricultural land may vary from USD 20.6 million to USD 102 million for a flood with an annual exceedance probability from 0.5 to 0.001. Moreover, the built-up area in 2040 will be exposed to flooding with an annual exceedance probability of 0.04 to 0.001 due to its distance from the river, resulting in damages of USD 30.5 million to USD 81.44 million, respectively. Considering the significant flooding in the area, even during 2- and 5-year return periods, future LULC transformations will lead to increased vulnerability in agricultural land and built-up areas, posing a significant flood risk. As shown in Figure 17, the EAD for built-up areas in 2021 was around USD 91 million, with projections indicating a rise to approximately USD 220 million by 2040. Similarly, the EAD for agricultural land with crops was estimated at USD 162 million in 2021, which could increase by roughly USD 376 million by 2040, given ongoing land use changes.

5. Discussion

The investigation of flood risk, encompassing vulnerability, hazard, and consequences, is crucial, especially in the context of notable events that occur repeatedly due to climate change and land use land cover (LULC) changes. Analyzing the impact of various flood risk variables is necessary to mitigate the loss of life and property damage and promote sustainable development [52]. This study employed multiple methods and approaches to assess flood risk in the Tajan watershed and examine the influence of dynamic human activities and complex geomorphological conditions on LULC changes.

5.1. LULC Modeling and Its Relative Impact

To facilitate socioeconomic development, our study tracked LULC changes using remote sensing, ARCGIS 10.5, and TerrSet software 2020, projecting shifts until 2040. Rigorous accuracy assessments, achieving over 80% overall accuracy and kappa index values, ensured the reliability of our predictions in the study area. The significant decline in forestland and rangeland from 2001 to 2021, attributed to expanding agricultural and built-up areas, aligns with socioeconomic trends in the study area. This result is consistent with other studies that focused on LULC changes in the northern part of Iran [53]. Various factors significantly influenced the LULC modeling process. Cramer’s V analysis highlighted that distance from the road and elevation were the primary factors shaping LULC transformations in the Tajan watershed. The transportation network, particularly road expansion, emerged as a critical element capable of transforming and altering the landscape, thereby escalating the potential for flood damage. Furthermore, road construction in the study area, intended to enhance connectivity between villages and cities, has directly influenced the expansion of residential areas [27]. Most villages and agricultural lands in the Tajan watershed have been established in relatively flat or downstream locations with lower elevations, making these areas more susceptible to floods and waterlogging. Consequently, floods significantly impact people’s lives, their properties, and the region’s infrastructure. The increase in agricultural land and deforestation, along with a reduction in the decrease in impermeable surfaces, has led to greater runoff intensity and velocity, aligning with studies on land cover conversion effects on runoff volume and speed [13,54].

The future LULC projection indicated an increase in agricultural land (+7.2%), a slight increase in the built-up area (+0.42%), notable declines in rangeland (−5.88%) and forestland (−2.3%), and minimal change in water bodies (+0.01%) over the past two decades. Rangeland will experience the most significant reduction compared to other land types due to increased human activities. This result indicates that the percentage of deforestation will be reduced due to the current strict rules for water management projects. Areas projected to become built up under the land use model must be considered as the most suitable areas for future development; however, these areas are prone to flooding and need to be controlled by a disaster management organization. These findings are consistent with previous studies on LULC’s future simulation of land use changes in rapidly urbanizing areas [55,56].

5.2. Flood Vulnerability Mapping Considering LULC Change Scenarios

This study utilized the Random Forest algorithm, an advanced ensemble model, to develop and validate a flood vulnerability map for Iran’s Tajan watershed. The Random Forest model, recognized as an effective classification tool, has shown superior performance in flood vulnerability modeling in previous studies [57,58,59]. Five key factors were identified as the most influential in determining flood vulnerability in the study area, with minor variations between past and future projections due to deforestation, human activities, road construction, and agricultural encroachment. In 2021, the distance from the river was identified as the most significant factor for flood susceptibility, followed by altitude, land use, slope, drainage density, and rainfall. Similarly, in the year 2040, the most important factors remained the same but with some degree of dynamism. The consistent significance of flood susceptibility factors from 2021 to 2040, with minor changes, reflects their inherent hydrological importance [60]. The LULC alteration scenarios utilized in this study were based on the hypothesis that the LULC transformation from 2021 to 2040 would follow the patterns observed in the previous 20 years (2001 to 2021). The results indicate that extreme LULC transformations are likely to occur at lower altitudes near the rivers. Consequently, the reduction in forestland and the increases in farmland and built-up areas at lower altitudes will amplify the intensity and frequency of floods in this region. These findings align with earlier research that links increased runoff velocity and flood risk to deforestation in downstream areas [61,62,63].

The vulnerability map was classified, using the break classification technique, into five categories: very low, low, moderate, high, and very high. The vulnerability map based on the LULC scenario for 2021 shows that the low-risk area (1047.75 km²) encompasses a major part of the study area. In addition, the high- and very high-risk classes comprise approximately 533.4 and 838.2 km² and are located in the downstream zone near the river. The map of predicted flood vulnerability considering the LULC transformations occurring by 2040 reveals that the very high- and high-flood-risk class areas would increase by about 746.76 and 883.92 km², respectively. However, the low- and very low-class areas will be reduced by about 708.66 and 662.94 km². Most of the downstream agricultural lands (rice fields) and built-up areas are located in high-risk and very high-risk flood zones and would be subject to serious flood financial and human losses. The high- and very high-flood-risk zones in both scenarios mainly cover the northern and western parts of the watershed in the vicinity of the main river and its branches on a lower slope. Low drainage density is one of the Tajan watershed’s characteristics, which is one of the main reasons for its high flood vulnerability. Conversely, the southern and eastern parts of the watershed enjoy lower vulnerability to flooding due to their higher elevation. The RF model was successfully validated for both periods, with AUC training and validation rates exceeding 0.90 in both the 2021 and 2040 scenarios. These results are consistent with previous studies’ findings and underscore the RF method’s reliability for accurately mapping flood susceptibility [64]. Several prior studies also support our findings regarding the significant impact of LULC changes on flood probability zones [65,66,67].

5.3. Flood Hazard and Damage Estimation Based on LULC Change Scenarios

This study integrates flood vulnerability and hazard data to assess the impact of flood consequences on different land use categories. The high-risk zone along the river, prone to flooding, was identified for evaluating damages on agricultural land and built-up areas under various LULC change scenarios. The HEC-RAS hydraulic model, calibrated with 30 years of flood discharge data, simulated inundation depth and intensity for different return periods (Q2 to Q1000). The model demonstrated accurate performance, aligning with similar studies [68,69]. For the 2-to-10-year return period, the average maximum inundation depth is 3.4 m, increasing to 10.63 m for higher return periods. In the low and medium return period floods, the difference in the maximum inundation depth is not extreme, probably because there is not a large difference in elevation between the river channel and adjacent areas. The results indicate that topography has a crucial impact on confining the inundated zone during extreme floods, where a further increase in the flood volume is converted into an increase in depth rather than area coverage [50]. In the studied watershed, sometimes, a high inundation depth occurred even beyond the river channel on the border between the river and agricultural land due to the formation of artificial levees and the shifting of the channel. Channel shifting in the Tajan River occurs due to human activity, loose soil, and subsequent erosion [70,71]. The depth–damage function was applied to estimate the direct financial cost of damages to the building and agricultural lands (with the crops) in various flood return periods under previous and future LULC change scenarios. A similar function has been used in other damage studies by Alipour et al. in 2020 and Narayan et al. 2017 at different locations [72,73].

Our results showed that due to the proximity to the river, the agricultural lands are exposed to all flood return periods (Q2 to the Q1000 year); however, the built-up areas a short distance from agricultural lands are susceptible to flood severity in the 25- to 1000-year return period. The results illustrate an increase in average damage between 2021 and 2040. Specifically, the average damage to the built-up area is projected to rise by USD 169.81 million, while the average damage to agricultural lands is expected to increase by USD 289.28 million. The expected annual flood damage (EAD) will increase in the future LULC scenario due to the limited support provided for the flood risk when making land expansion plans. The expected annual damages (EADs) for the built-up area will increase by USD 129 million from 2021 to 2040. Similarly, for agricultural land (with crops), the EAD is expected to rise by USD 214 million, reflecting the ongoing LULC changes. Damage analysis using a depth–damage curve is a standard damage calculation method; however, due to some uncertainties, this method may overestimate flood losses [51]. The findings of this study highlight that the Tajan watershed will face increased flood risk if risk orientation is not considered in future land use planning to mitigate damage and ensure long-term sustainability, so it is crucial to prioritize the implementation of flood risk mitigation strategies. The spatial data on flood risk and land use changes provided in this research can assist stakeholders, including the Ministry of Land, in identifying areas that require land use policy intervention. Additionally, the developed framework can be valuable for land use planners to evaluate the impacts of land use transformations on flood risk and its consequences. This research represents an initial effort to quantify the losses resulting from potential land use transformations and the associated flood risk in the Tajan watershed. While numerous studies have focused on either modeling land use transformation or flood risk, there is a scarcity of assessments that consider their combined impact.

In general, this study aims to reduce uncertainty through the integration of various techniques and models, including historical trend analysis for predicting land use and land cover (LULC) changes, rigorous model evaluation, and standardized data processing and visual analytics. However, it is important to acknowledge that the complete elimination of uncertainty is not achievable. For further assessment, exploring interactions between climate change scenarios—considering greenhouse gas emissions and socioeconomic factors—and their combined impact on flood risk with LULC changes is recommended. Additionally, incorporating a broader range of qualitative factors such as social vulnerability and community resilience into the analysis will contribute to a more comprehensive understanding of flood risk. The utilization of optimized models can further improve the accuracy of projections. The transparent communication of uncertainties and careful interpretation of findings will aid informed decision-making processes.

It is also worth noting that the 2040 damage projections in this study are based on current infrastructure and technology for flood control. This study applied consistent inflation rates and assumed annual GDP growth rates of 4.5% for built-up areas and 5.5% for agricultural land. However, we acknowledge that future factors, such as fluctuations in currency values, unexpected economic shifts, or changes in growth rates, could influence these estimates. To strengthen future projections, it would be beneficial to consider a range of economic scenarios that account for these potential variables.

6. Conclusions

This study developed a flood risk assessment framework that integrates flood vulnerability, hazard, and flood consequences with the spatiotemporal changes in land use and land cover (LULC) scenarios in the Tajan watershed in Iran. We applied a series of computational operations and complex geoprocessing techniques to identify high-risk areas, measure flood depths, and calculate damage influenced by land use changes. The primary findings are as follows:

(i) LULC Transformation: This study identified rapid LULC changes, particularly increases in agricultural and built-up areas, especially near rivers and downstream regions, while forest and rangeland areas decreased. This shift, confirmed by spatial analysis, is anticipated to escalate runoff and flood occurrences, especially downstream. These results underscore the importance of considering flood impacts when planning land expansion to reduce future flood risks.

(ii) Flood Vulnerability Mapping: Random Forest models effectively mapped flood vulnerability under LULC changes, demonstrating high AUC values and accurately identifying the middle and downstream regions of the Tajan watershed as highly vulnerable. This vulnerability is driven by factors such as lower elevations, gentle slopes, increased rainfall, and notable LULC transformations. The study projects a 43% rise in high-vulnerability areas by 2040, primarily due to human activities and ongoing LULC changes. Key contributors to flood risk include proximity to the river, elevation, land use type, slope, drainage density, and rainfall, each varying in influence across different scenarios.

(iii) Economic Impact Assessment: Using HEC-RAS and global damage functions, this study assessed the financial impacts of LULC-induced floods under two scenarios. Direct losses were calculated for buildings and agricultural land in high-risk flood zones across various return periods. The analysis showed that flood impacts on agriculture were largely influenced by flood height and duration, while built-up areas experienced significant damage in events with a return period of 25 years or more. Due to LULC changes, damages are projected to increase, indicating growing exposure to flood risks. The estimated annual flood damage (EAD) for agriculture is set to rise from USD 136.85 million in 2021 to USD 376.10 million in the 2040 scenario. Similarly, EAD for built-up areas is expected to increase from USD 91.43 million in 2021 to USD 220.49 million in the 2040 scenario. These finding underscores the critical role of land use planning in reducing flood-associated risks.

This study highlights flood risk zones resulting from future land cover changes, aiding land development decision-makers in determining appropriate flood conservation and damage mitigation actions. The projected increase in flood damage, primarily due to agricultural and built-up area development in floodplains, allows planning authorities to employ local planning instruments. These measures can redirect land development and promote structural adaptation strategies for agricultural land and buildings in areas prone to potential flood hazards, thereby reducing damage potential.