Machine Learning Algorithms like SVM, RF, Decision Tree, etc., have turned out to be efficient methods for research in today’s date due to their impeccable accuracy and reliability. Support Vector Machine (SVM) is a type of supervised machine learning that can effectively identify intricate patterns in noisy and complex datasets, and due to their simplicity and adaptability, they can achieve balanced predictive accuracy even in situations where there are limited samples (Hongmao, 2016). Random Forests (RF) improve prediction accuracy and efficiency by randomly selecting features for each decision split, reducing correlation between trees, and increasing the diversity of the model(Breiman, 2001). Decision Tree (DT) is an inductive algorithm used for classification and prediction, where classification rules are represented as decision trees derived from a set of disorderly and irregular instances, and the tree is constructed in a top-down recursive manner by comparing attributes between internal nodes and making decisions based on different attributes, ultimately leading to a conclusion at the leaf nodes (Dai et al., 2016). The significant advancements in machine learning and artificial intelligence, including logistic regression, decision trees, artificial neural networks, random forests, and support vector machines, have gained immense importance due to their ability to handle large datasets and deliver high levels of accuracy (Ruidas et al., 2021).

Several researchers have used ML algorithms to create remarkable research projects in variety of sectors such as (Ruidas et al., 2022) in Hydrogeochemical Evaluation of Groundwater Aquifers (Ruidas et al., 2022), in water resources vulnerability assessment (Ruidas et al., 2022), in flood-susceptibility assessment (Ruidas et al., 2021), in Characterization of groundwater potential zones (Jaydhar et al., 2022), in Hydrogeochemical evaluation and health risk from arsenic and fluoride, etc.

The use of ML algorithms in disaster prediction, Vulnerability and risk assessment, mitigation has become a sought-after procedure in the current scenario and this research has used the same to generate comprehensive risk and vulnerability zones of the fragile Bhitarkanika National Park region. Pham et al. (2019) in their research used hybrid machine learning models, including bagging (BA), random subspace (RS), and rotation forest (RF), with alternating decision tree (ADTree) as base classifier for the spatial prediction of Landslides (Parvin et al., 2022). in their study used 3 ML models, namely, Bayesian logistic regression (BLR), the artificial neural networks (ANN), and the deep learning neural networks (DLNNs) for flood vulnerability assessment in a densely urbanized city. Opella and Hernandez (2019) in their study generated flood susceptibility and probability map using SVM and obtained a robust flood map that clearly outperforms the traditional methods. Xiong et al. (2019) in their study adopted SVM model for flash flood vulnerability assessment and mapping in China.This study uses the SVM-RBF model to generate a robust vulnerability and risk Map of the Bhitarkanika National Park region taking in view previous studies, which have used the same model for its impeccable accuracy. The map generated using SVM-RBF exhibits an accuracy of 99.54% with a complementing Kappa Index of 99.18% compared to the 91.12% accuracy using traditional AHP, thus solidifying the SVM-RBF model as a formidable classification ML classification.

Pre-processing of data, such as satellite imagery and digital elevation models (DEM), is crucial for the processing, analysis, and modeling in this study. In order to map the land use and land cover (LULC) of Bhitarkanika National Park, satellite imagery underwent pre-processing steps including band stacking, clipping, mosaicking, and normalization using the min-max scaler. These pre-processing tasks were performed using QGIS. Similarly, the SRTM DEM was pre-processed using both ArcGIS and QGIS. The DEM was initially clipped to the study area by applying a mask. Auto co-registration and filling techniques were then employed using the hydrology toolbox in ArcGIS and QGIS to ensure alignment with the LULC data and to address sinkholes, which are often not captured by satellites. The DEM was further reclassified to generate an elevation map, and the slope was calculated using the Arc Toolbox.

These pre-processing steps were undertaken to ensure the data was appropriately prepared for subsequent analysis and modeling in the study.

Land use and land cover change have become central to current strategies for managing natural resources and monitoring environmental changes(Kaul and Sopan, 2012). The standard land use and land cover (LULC) classes (Table 2) were selected based on the literature review and local LULC classification scheme. Based on the previous literature it was observed that uniform LULC classification scheme is missing in disaster study (Hao et al., 2022). Landsat 8 Operational Land Imager (OLI) and Landsat 5 Thematic Mapper (TM) images of 2021 and 2000, respectively, with cloud cover of less than 2%, were obtained from the google earth engine using JavaScript codes. Further processing was done in QGIS, including feature extraction(e.g., NDVI, NDBI, etc.), classification, post-processing, accuracy assessment, and change analysis.

The Normalized difference Vegetation Index (NDVI) is widely used in classifying land use/cover which was calculated using following formula (Ruidas et al., 2021): N D V I = N I R − R E D N I R + R E D (1)

The value of the NDVI varies in between +1 and −1. NDVI is equal to +1 shows healthy vegetation while −1 shows waterbodies. In addition, Normalized Difference Built-up Index provide vivid information of the built-up which was calculated using following formula (He et al., 2010). N D B I = B a n d 5 − B a n d 4 B a n d 5 + B a n d 4 (2)

Higher the value of NDVI shows more the built-up information which lower values shows vegetation and other land use classes.

Table 2 shows the training and test samples used for the training and validation of the classification model.

There are various types of classifiers in machine learning (ML). This study uses the SVM classifier with Radial Basis Function (SVM-RBF) via the OTB toolbox to generate LULC maps and the flood risk map (Deroliya et al., 2022). This is because machine learning algorithms outperform any complex decision-making compared to other traditional algorithms(Farhadi and Najafzadeh, 2021; Deroliya et al., 2022). This study selected SVM-RBF because this algorithm is robust for complex problems compared to the other machine learning algorithms such as Random Forest, Decision Tree, etc (Ruidas et al., 2021; Ruidas et al., 2022). Two LULC maps (e.g., 2021 and 2000) were generated using the image classifier function in the OTB tool. The accuracy assessment of all the maps generated using ML was done by computing their confusion matrix using the OTB tool. Change detection analysis, one of the import approaches, is incorporated with the flood risk analysis (Gharagozlou et al., 2011; Lawal et al., 2014) carried out between the 2000 and 2021 LULC maps using the post-processing algorithm and Raster Unique Values Report function in QGIS.

When floods hit inhabited areas, significant losses are usually registered in terms of both impacts on people (i.e., fatalities and injuries) and economic impacts on urban areas, commercial and productive sites, infrastructures, and agriculture. To properly assess these, several parameters are needed, among which flood depth is one of the most important as it governs the models used to compute damages in economic terms(Cian et al., 2018). In this study, the Raster calculator was used in ArcGIS to analyze flood/water depth manually. The inundation depth is estimated to be 2.5 m through multiple literature reviews and field surveys. Based on the last 20 years’ flood inundation information, the binary mask was created as one and none flooded area as zero. O u t p u t   R a s t e r = D E M   x   B i n a r y   M a s k (3)

The result gave the elevation values of DEM for areas, which are flooded, and zeros for non-flooded areas. Consequently, the highest elevation value represents the water table. W a t e r   D e p t h = V a l u e _ w a t e r _ t a b l e   x   B i n a r y   M a s k (4) W h e r e , v a l u e _ w a t e r _ t a b l e = 2.5 m

The resultant raster thus obtained represents the Water/Flood depth of the study area.

The shortest straight-line distance connects all sites or the Euclidean distance (Zhang, 2019). Geoprocessing analysis is performed to fill sinks (pits) and to generate data on flow direction, flow accumulation, catchments, streams, stream segments, and watersheds. These data are then used to develop a vector representation of catchments and drainage lines from selected points that can then be used in network analysis (Soni, 2012). To calculate the Euclidean distance from the river, the process begins with stream delineation using the hydrology toolbox in ArcGIS. This involves using the fill tool followed by the flow direction tool, which determines the downslope direction of each cell and helps identify the flow paths of the streams. The flow accumulation tool is then applied to estimate cumulative flow, representing the total weight of cells flowing into each downslope cell. By setting a threshold value, the number of streams included in the final layer can be controlled. Lower threshold values result in more streams, while higher values reduce the number of streams.

Once the streams are delineated, the Euclidean distance tool is used to calculate the distance from the stream. Similarly, the coastline of BNP is manually digitized, and the Euclidean distance tool is applied to determine the distance from the coast. These steps enable the calculation of the Euclidean distance from both the river and the coastline, providing valuable information for further analysis and modeling in the study.

The goal of flood hazard assessment is to understand the probability that a flood of a particular intensity will occur over an extended period of time. Hazard assessment aims to estimate this probability over periods of years to decades to support risk management activities(Wright, 2015). Intensity is typically defined as the sum of flood depth and horizontal flood extent. However, depending on the circumstance, other intensity parameters like flow velocity and flood duration may also be significant (Stefanidis and Stathis, 2013; Farhadi and Najafzadeh, 2021; Deroliya et al., 2022). Hydrological models like water depth and other factors like frequency and area of Impact were used to estimate the flood hazard. H S = F S   x   A I S   x   I S (5) Where, HS = Hazard Score FS = Frequency Score AIS = Area of Impact Score IS = Intensity Score

Aside from flood danger, another critical factor in flood risk is flood vulnerability. Understanding a system’s vulnerability will help you predict how floods may damage it. Examples of potential systems include physical structures like homes or bridges that might sustain damage or destruction, a company or service whose supply chain might be disrupted, or a community that might experience fatalities, property losses, and detrimental health effects following a flood(Wright, 2015). Various indicators are used to estimate the vulnerability of BNP. The indicators used are elevation, slope, water depth, distance from the coast, and distance from the river (Stefanidis and Stathis, 2013; Farhadi and Najafzadeh, 2021). The indicators have been reclassified into ‘Cost’ and ‘Benefit’ to better understand the assessment (Table 3).

Further, all the cost and benefits indicators are normalised, and the AHP model was applied to achieve flood vulnerability of BNP using formula 6-7 in the raster calculator tool in ArcGIS.

1. Normalisation

The practice of making specific data that are separated by time periods identical, such as atmospheric correction or pixel resampling, so that an acceptable change may be observed without being impacted by other factors is referred to as normalisation.

Cost indicator: N o r m a l i s a t i o n = 1 − I n d i c a t o r − i n d i c a t o r _ m i n i n d i c a t o r _ max ⁡ ⁡ –   i n d i c a t o r _ m i n (6)

Benefit Indicator: N o r m a l i s a t i o n = I n d i c a t o r − i n d i c a t o r _ m i n i n d i c a t o r _ m a x   –   i n d i c a t o r _ m i n (7)

2. Analytic Hierarchy Process- Weight Overlay Analysis

Analytic Hierarchy Process (AHP) is a robust multi-criteria decision-making (MCDM) was used to achieve the weight of the factors for the overall decision-making (Ouma and Tateishi, 2014; Rahmati et al., 2016; Kumar et al., 2021; Parsian et al., 2021). This model is widely used in raster-based GIS overlay analysis in several applications such as land suitability analysis, flood risk and vulnerability analysis, zoning, and site suitability analysis (Mustak et al., 2018). In AHP, the following sub-processes were employed to derive the weight of the indicators (Stefanidis and Stathis, 2013), e.g., 1) selection of indicators and arrange in the square-matrix, 2) indicators were compared, and relative importance given based on the Saaty’s nine-point scale of absolute number, 3) normalized weight of the individual indicator was derived using the geometric mean method (Table 4). The weighted overlay analysis is one of the most used methods to address multi-criteria issues like site selection, land suitability analysis, and assessing model appropriateness (Kumar et al., 2021).

The sum of the normalized weight is 1. After the calculation of weights, raster calculator was used to derive the final vulnerability index VI), which varies from 0 to 1 using the following Formula 6. The VI, equal to 0, shows low vulnerability, while 1 shows high vulnerability. V u l n e r a b i l i t y = F 1   x   0.3 + F 2   x   0.3 + F 3   x   0.15 + F 4   x   0.15 + F 5   x   0.1 (8)

The final output raster is symbolised using the quartile method, and the resultant raster is the vulnerability map.

The most common approach to define flood risk is that it is the product of hazard, i.e., the physical and statistical aspects of the actual flooding (e.g., the return period of the flood, extent, and depth of inundation, and flow velocity), and the vulnerability, i.e., the exposure of people and assets to floods and the susceptibility of the elements at risk to suffer from flood damage(Serda et al., 2002). After calculating the Flood Hazard and flood vulnerability, it becomes relatively simpler to calculate the Flood Risk. F l o o d   R i s k = F l o o d   H a z a r d   x   F l o o d   V u l n e r a b i l i t y (9)

The hazard and vulnerability maps produced before are normalised first and then multiplied using the raster calculator tool. The resultant raster gives us the flood risk map of BNP, which is further classified into High, Medium, and low based on the corresponding intensity value.

As observed from the resultant map, a substantial area of BNP has a water depth of 0 m constituting to 383.92 sq. km and 78% of the total area (Figure 4). This area represents minimal flood hazard. An area of 47 sq. km or 9.5% of the total area has a water depth of 1 m representing low flood hazard. A water depth of 1.5 m is observed across 5.18 sq. km or 1.05% of the area representing a medium flood hazard. The Brahmani and Dharma river systems, as well as the areas of Ramchandrapur, Jagannathpur, Padmanavpur, Narayanpur, Saradaprasad, Paramanandpur, and Mohanpur, exhibit a water depth of more than 2.5 m covering an area of 55.14 sq. km or 11.23% of the total area (Table 6). These areas are most prone to flooding and have a high flood hazard.

BNP has a maximum elevation of 23 m. The lower the elevation higher is the risk of getting affected by flood and vice versa. As observed in the elevation map, all the areas under mangrove vegetation have a medium to high elevation, which makes these areas resilient to flooding (Figure 5; Table 7). The areas surrounding the mangroves and the riverbanks subsequently have lower elevations. The areas near the coast also have a lower elevation, excluding those covered by mangroves.

The slope is the most crucial aspect of hydrology since it directly affects surface runoff and floods. Since low-elevation locations often have a gentle or low-level slope (0-1.39 m), they are more susceptible to flooding and waterlogging because steep slopes generate more incredible velocity than flat or gentle slopes and may dispose of runoff more quickly. Runoff from a level or gently sloping land is collected and released gradually. In contrast to high-gradient slopes, low-gradient slopes at lower reaches are more susceptible to flooding (Ramesh and Iqbal, 2022).

The BNP area has varied slope values and is unevenly distributed (Figure 6; Table 8). A large area of the national park appears to have low-medium slope values (1.39-2.09°). High slope values (2.09-19.72 m) appear to be scarce and scattered.

The result was obtained using the Euclidean distance tool in ArcGIS. The BNP area is bordered by three rivers: Brahmani, Baitarani, and Dharma. The Dharma River is formed at the confluence of the Brahmani and Baitarani Rivers. The Brahmani River covers a significant portion of the riverine area within BNP. It both surrounds and cuts through the national park, eventually flowing into the Bay of Bengal. As a result, many areas in BNP are located near the riverbanks and are susceptible to flooding. The geography of BNP is characterized by its surrounded by rivers and the ocean on all sides.

.The map produced is classified into three classes: High Proximity (<956.4 m), Medium Proximity (956.40-2646.30 m), and Low Proximity (>2646.30 m) (Figure 7; Table 9). The map displays the surrounding areas of Praharajpur, Gobardhanpur, Raj Nagar, Ramchandrapur, Govindpur, Subarnapur, as well as the central villages of Balabhdrapur, Purushottampur, Gupti, Padmanavpur, Jaganaathpur, and others. These areas are situated along the riverbanks of the Brahmani and Dharma rivers, making them highly proximate to these water bodies.

The Bay of Bengal lines the whole eastern area of BNP. This Proximity to the sea makes the coastal areas of BNP extremely vulnerable to tidal floods, especially during storm surges and tsunamis. Added to that, the Bay of Bengal is very prone to cyclones. The Gahirmatha Marine Sanctuary is shielded from such degradation due to the presence of mangroves. The map produced is classified into three classes: High Proximity (<4418.60 m), Medium Proximity(4418.60-8764.90 m), and Low Proximity (>8764.90 m). The coastal areas of Paramanandapur, Karanjia, Kanhupur, Gupti, Barunei, Satabhaya, Pentha, Jamboo, Batighar, Suniti, Kansarbadadandua, Ramanagar, and Baulakani all fall under the high proximity class and are highly vulnerable to tidal floods (Figure 8; Table 10).