Optimizing ensemble learning for satellite-based multi-hazard monitoring and susceptibility assessment of landslides, land subsidence, floods, and wildfires

Abstract

The preparation of accurate multi-hazard susceptibility maps is essential to effective disaster risk management. Past studies have relied mainly on traditional machine learning models, but these models do not perform well for complex spatial patterns. To address this gap, this study uses two meta-heuristic algorithms (Genetic Algorithm (GA) and Particle Swarm Optimization (PSO)) to provide an optimized Random Forest (RF) model with better predictive ability. We focus on four significant hazards—landslides, land subsidence, wildfires, and floods—in Kurdistan Province, Iran, using Sentinel-1 and Sentinel-2 satellite imagery collected between 2015 and 2022. Furthermore, two models of RF-GA and RF-PSO were utilized to create multi-hazard susceptibility, which were evaluated using receiver operating characteristic (ROC) curves and area under the curve (AUC). The RF-GA algorithm achieved 91.1% accuracy for flood hazards, 83.8% for wildfires, and 99.1% for landslide hazards. In contrast, utilizing RF-PSO resulted in a 95.9% accuracy for land subsidence hazards. The combined RF-GA algorithm demonstrated superior accuracy to individual RF modeling techniques. Furthermore, eastern regions are more prone to floods and land subsidence, whereas western areas face more significant risks from landslides and wildfires. Additionally, floods and land subsidence exhibit a considerable correlation, impacting each other’s occurrence, while wildfires and landslides demonstrate interacting dynamics, influencing each other’s likelihood of occurrence.

Introduction

According to the World Health Organization’s (WHO) classification, disasters are classified as natural, human-made, and complex disaster¹. Natural hazards are phenomena of natural origin that, upon occurrence, can lead to death, injury, or other adverse health effects, damage to property, livelihoods, and services, as well as societal and economic disruption or environmental harm^2,3. Given the increasing incidence of these hazards and the resulting losses, conducting investigations, monitoring, assessment, and planning to mitigate their risks are paramount^4,5.

A World Bank report on major centers of natural hazards found that around 3.8 million km² and 790 million people are relatively exposed to at least two hazards. In comparison, about 0.5 million km² and 105 million people are exposed to three or more hazards. Despite this, natural hazards are typically studied independently^6,7. Examining natural hazards in isolation can result in inaccuracies in assessing the danger or likelihood of particular extreme natural events occurring in those areas^8,9. Given these conditions, individual assessments of natural hazards may not adequately address the need for planning and risk reduction¹⁰. Since most places are based on numerous interrelated hazards that generate cascade effects, an integrated approach is required for improved hazard management in every region^11,12.

Multi-hazards refer to areas where hazardous events may co-occur, cascading or accumulate over time, considering their potential interrelated effects^8,13. The frequency, duration, distribution, or severity of a process can worsen when efforts to mitigate one hazard interact with other processes¹⁴. These factors underscore the importance of studying hazards in a multifaceted manner. Multi-hazard risk assessment can help control hazard interactions^8,15. Therefore, effectively reducing multi-hazards requires analyzing individual risks and their mutual influences. In natural resource management and development planning, understanding areas susceptible to various hazards is crucial^16,17.

The United Nations has launched a worldwide initiative to improve disaster management and risk reduction in the face of multiple hazards within the framework of Agenda 21 and sustainable development¹⁸. Guideline 21 advocates for the “assessment of multi-hazards” as part of risk management and planning for human settlements in areas susceptible to these hazards¹⁹. In supporting socio-economic development and enabling sustainable natural resource management, policymakers need knowledge of regions experiencing multi-hazards, which is directly tied to complete sustainable development²⁰. This underscores the importance of integrating multi-hazard assessment into the broader framework of sustainable community management^19,20.

To mitigate human casualties and reduce economic losses, natural hazard management is essential, necessitating the assessment of multi-hazard situations and identifying priority areas for sustainable management⁷. Consequently, it is necessary to conduct further research to achieve prevention and mitigation solutions by delineating susceptible areas for hazard occurrence through zoning hazardous regions^21,22. Susceptible areas and those with a high potential for hazard occurrence can be identified through hazard zoning, allowing for the provision of solutions and the application of appropriate control and management methods to prevent or mitigate hazards to a certain extent, thereby reducing resulting damages^7,23,24.

Analytical techniques, crucial in addressing complex spatial problems, involve investigating spatial characteristics and relationships²⁵. These techniques aim to acquire valuable information and respond to spatial questions^26,27. Geographic Information System (GIS) facilitates spatial modeling for phenomena, including natural hazards, aiding mapping and analysis^27–29. In spatial modeling, considering the probability of the cumulative effects of multi-hazards enhances the modeling of various individual risks^30,31. A spatial approach in multi-hazard assessments examines vital information about hazard-prone areas, vulnerable regions, and their geographical distributions and infrastructures^32–34.

GIS modeling uses remote sensing data to identify and monitor natural phenomena with increased speed and accuracy^35,36. This results in enhanced spatial resolution, computational capabilities, quantitative precision, and data sharing for the general public³⁷. Remote sensing is a crucial tool for monitoring changes in the Earth’s surface and is increasingly used as a foundation for early warnings of hazardous events³⁸. Satellite images were utilized in this study due to their advantages, such as large-scale examination, suitable spatial resolution, and accessibility for various timeframes, for monitoring each hazard’s occurrence³⁹.

Due to its unique climatic conditions, topography, and geology, Iran is prone to frequent natural disasters, often accompanied by severe damage^40,41. Natural hazards such as floods, wildfires, land subsidence, and landslides are among the most prevalent risks in Iran, causing severe impacts annually, with their frequency rising due to changes in climate conditions, land-use alterations, and development^42–44. Natural disasters like floods, wildfires, land subsidence, and landslides are common in Iran’s northwest, which includes Kurdistan province⁴⁵. Therefore, this study aims to monitor four hazards: landslides, land subsidence, floods, and wildfires, and to produce susceptibility maps for these hazards individually and in combination in Kurdistan province, Iran.

Numerous studies have investigated multi-hazard risks across various regions. Bathrellos⁴⁶ utilized the analytical hierarchy process (AHP) to map flood, landslide, and earthquake risks in northeastern Greece, identifying low-hazard areas in the southern region. Similarly, Sheikh¹¹ employed a technique for order preference by similarity to ideal solution (TOPSIS) and maximum entropy methods to assess flood, landslide, and soil erosion risks in Golestan Province, identifying hazardous areas like Minudasht and Ramian. Kaur⁴⁷ applied AHP to evaluate landslide and earthquake risks in India, finding that most residential buildings were in low-risk zones. Pourghasemi⁴² combined an adaptive network-based fuzzy inference system (ANFIS) and grey wolf optimizer (GWO) algorithm to map flood, landslide, and earthquake risks in Lorestan Province with higher accuracy in flood modeling.

Skilodimou⁴⁸ used the AHP method to map the risks of landslides, floods, and earthquakes in Greece, noting high-risk areas in the western and northeastern regions. Rahmati⁴⁹ employed machine learning algorithms for avalanche, rockfall, and flood risk mapping in Iran, with support vector machine (SVM) showing higher accuracy. Pourghasemi⁵⁰ utilized machine learning to model multi-hazard risk maps for floods, landslides, and forest fires in Fars Province, revealing that 52.3% of the area was exposed to at least one hazard. Yanar⁴³ employed Fuzzy Mamdani and AHP methods for modeling landslides and floods in Turkey, finding residential areas at multi-hazard risk.

Cao⁵¹ used gradient boosting to map China’s landslides, rockfall, and debris flow, identifying susceptible regions mainly in the southeast and middle of Jiuzhaigou. Aksha³² utilized the AHP method to map landslide, flood, and earthquake risks in Nepal, identifying high-risk areas in eastern Daran along the Suti River and southwestern Daran on the left bank of the Sardu River. Nachappa¹⁹ employed random forest (RF) and SVM algorithms to assess flood and landslide risks in Salzburg, Austria. RF algorithm demonstrated higher accuracy in multi-hazard risk mapping, particularly highlighting the elevated probability of flooding. Bordbar⁵² investigated flood, landslide, and earthquake hazards in Kermanshah Province, Iran, using metaheuristic and machine learning algorithms. Ullah⁵³ developed a susceptibility map for floods and landslides using a deep-learning approach. Akbar⁵⁴ examined landslides, floods, and avalanches in India using the statistical frequency ratio (FR) method. Pourghasemi⁵⁵ employed combined machine-learning methods to study floods, landslides, forest fires, and earthquakes.

Researchers have recently utilized machine learning models in the spatial modeling of natural hazards^42,56. The advantages of these models include their lack of simple statistical assumptions, the ability to examine non-linear and complex relationships between data, and the ability to manage and control data limitations^57,58. RF, a machine learning algorithm, is often used for its ease of use and frequently provides excellent results for classification and regression tasks⁵⁹. The primary limitation of RF is that many trees can slow down the algorithm for real-world predictions and make it less efficient^59,60. To enhance the performance of the RF algorithm and improve its accuracy and speed, various feature evaluation and combination techniques, along with metaheuristic algorithms, can be employed⁶¹.

Metaheuristic algorithms provide various advantages for optimizing the RF algorithm. These algorithms excel at searching the global solution space, allowing for finding optimal or near-optimal RF hyperparameter values. Given RF’s efficacy in handling high-dimensional data, the problem is determining the ideal number of trees, tree depth, and other hyperparameters⁶². Metaheuristic algorithms search these hyperparameters effectively, identifying the optimal combination to maximize RF performance. Metaheuristic algorithms and RF have been employed to model the spatial distribution of natural hazards such as floods³⁹dust storms⁶³and landslides⁶⁴. However, this combination has yet to be utilized for spatial modeling multi-hazards.

Despite significant progress in multi-hazard risk assessments, current studies have several limitations. First, many studies utilize traditional machine learning models that are not fully optimized for predictive performance and, thus, cannot be applied in complex spatial hazard scenarios. Second, although RF analyses are widely used to model multi-hazards, the vast potential of RF has not been fully realized due to the lack of rigorous hyperparameter tuning with specifically metaheuristic optimization. Third, land subsidence, an important geohazard rapidly developing, is omitted from previous multi-hazard frameworks, focusing only on floods, landslides, and wildfires. Finally, many studies depend on official hazard inventories or static datasets, which may lack the spatial and temporal resolution necessary for accurate susceptibility mapping.

To fill these gaps, this research introduces a new spatial modeling and susceptibility mapping framework of four significant natural hazards (floods, wildfires, landslides, and land subsidence) across Kurdistan Province, Iran. As part of this research, we introduce several important innovations. First, we develop hazard-specific optimization of the RF algorithm using two metaheuristic techniques - Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) - representing the first interdisciplinary use of these algorithms in a multi-hazard context to improve the prediction accuracy. GA is an evolutionary algorithm inspired by biological evolution and applies concepts of selection, crossover, and mutation to evolve a population toward an optimal solution gradually⁶⁵. GA is highly employable in discrete and mixed-integer optimization applications, especially regarding tuning hyperparameters in RF models. PSO is simple to implement and known for rapid convergence, as the collective behavior of birds in flocks, or other social creatures, inspires it⁶⁶. Finally, the PSO uses a population of candidate solutions (i.e., particles) that simultaneously change all particle candidates’ positions in the search space relative to the personal best and global best positions to quickly optimize and update the continuous parameters⁶⁷. GA and PSO in this study provide a dual-optimization framework, where each algorithm offers unique advantages. GA’s genetic diversity helps avoid premature convergence, while PSO’s velocity-based updates ensure rapid convergence to optimal regions^65,66. Second, the research utilizes historical satellite imagery for direct hazard monitoring, rather than relying on aggregated or secondary hazard data. Specifically, Sentinel-1 imagery was employed for flood and land subsidence modeling (with InSAR (Interferometric Synthetic Aperture Radar) techniques), Sentinel-2 for wildfire detection, and Sentinel-1 data for modeling landslides based on ground-deformation signals. Finally, this research is among the first studies for the region to incorporate land subsidence into the multi-hazard susceptibility model, allowing for a more comprehensive and realistic spatial risk assessment.

Materials and methods

Research methodology

Figure 1 shows the six primary phases comprising the execution of this study: data collection, modeling, and evaluation. In the first step, the distribution maps of each hazard were processed and prepared using various satellite images, including Sentinel-1 and Sentinel-2. In the second step, satellite images and GIS analyses derived 19 spatial factors influencing each hazard. Pre-processing of raw data was conducted in the third step, including multicollinearity testing, determination of weight coefficients using the FR method, and identification of the importance of effective factors using the RRelif method. Spatial modeling of multi-hazards was performed in the fourth step by combining the RF algorithm with two metaheuristic algorithms, GA and PSO, in the Python programming environment. In the fifth step, susceptibility maps for each of the four hazards were prepared using the developed models. Finally, the multi-hazard risk map in the study area was generated. The susceptibility maps and modeling were evaluated using various indices in the final step.

Fig. 1.

General flowchart of the research.

Study area

The province of Kurdistan, centered on the city of Sanandaj in western Iran, spans an area of over 29000 Km²constituting 7.1% of the country’s total area (Fig. 2). Geographically, the province is situated between 34°45’ to 36°28’ north latitude and 42°31’ to 48°16’ east longitude. The population of Kurdistan province is 2,152,000, with 66% residing in urban areas and the remaining in rural regions. Kurdistan is physically and climatically defined by its hilly landscape, which features lofty plateaus and wide valleys. The province has an approximate 2400 m variance in elevation between its highest and lowest locations. Warm and humid air masses from the Mediterranean impact the climate of Kurdistan, causing springtime scattered rains and winter snowfall. There are roughly 109 frost days annually in the area, and under typical weather conditions, it rains about 500 mm annually on average.

Fig. 2.

Study area’s geographical position in Iran and worldwide (Maps are made using ArcGis10.8 (version: 10.8; URL link: https://www.esri.com)).

According to the Iranian Landslide Database records, Kurdistan province ranks third in landslide frequency. Due to prolonged droughts in Iran and unregulated groundwater extraction, 18 provinces, including Kurdistan, are at risk of land subsidence. Kurdistan province is also highly susceptible to wildfires, with over 374,000 hectares of its 2,937,000 hectares comprising forests. Statistics reveal 1,124 wildfires and the burning of 7,364 hectares of forests and pastures in Kurdistan province between 2006 and 2015. In 2019, 271 incidents of wildfires occurred in the province’s rangelands and forests, resulting in the burning of 2,658 hectares. Kurdistan province witnessed significant damage in 2019 due to heavy rainfall, resulting in flooding. The floods caused damages of 2.2 billion dollars to infrastructure and agricultural sectors in the province. Moreover, the floods affected 999 residential units, of which 876 were demolished and required complete reconstruction. Additionally, 17 million dollars in damages were inflicted on the road infrastructure in three sectors, including roads, bridges, and road-related structures in the province.

Hazard detection by satellite imagery

In this study, Sentinel-1 and Sentinel-2 satellite images were utilized to generate inventory maps of flood, landslide, wildfire, and land subsidence occurrences, each of which is elaborated upon below.

Flood inventory map

For monitoring past flood events, Sentinel-1 images from the Google Earth Engine (GEE) system (https://earthengine.google.com/) were employed. The Ground Range Detected (GRD) Level 1 product was used for pre- and post-flood images from 2018 to 2022. A descending orbit was employed for this objective using a pixel size of 30 × 30 m and VH polarization. The preprocessing steps applied to the GRD data included: (1) updating orbit metadata using the apply-orbit-file function to refine satellite positioning; (2) border noise removal to eliminate low-intensity noise and invalid edge data; (3) thermal noise removal to correct for additive sensor noise in sub-swaths; (4) radiometric calibration to convert raw digital numbers into physically meaningful backscatter intensity values (σ°) using internal calibration parameters; and (5) terrain correction using the Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) to orthorectify the imagery and ensure accurate geolocation. Finally, the calibrated backscatter coefficient was converted to the decibel (dB) scale to enhance contrast and interpretability. After these corrections, a speckle filter was applied to the mosaic images before and after the flood. In the next step, change detection between these two series of images was performed, and a difference layer was generated. Every value above 1.25 was given a value of 1 and every value below 1.25 was given a value of 0 according to a pre-defined threshold. This procedure generates a binary raster layer that shows the possible flood extent. Also, multiple extra datasets were utilized to remove flood layer false positives. A global surface water dataset was utilized to remove permanent surface water areas from the flood layer. To exclude areas with a slope greater than 5%, a DEM based on SRTM data with a pixel size of 30 × 30 m was employed. Additionally, pixels with less than four neighbors in the flood layer were removed. Subsequently, for modeling and producing a flood susceptibility map (FSM), the potential flood locations generated using remote sensing images were converted to points. For this purpose, 762 flood points were obtained for modeling and validation (Fig. 3a).

Fig. 3.

Distribution map of hazards in the study area: (a) Flood, (b) Wildfire, (c) Landslide, and (d) Land subsidence (Maps are made using ArcGis10.8 (version: 10.8; URL link: https://www.esri.com)).

Wildfire inventory map

Sentinel-2 images in the GEE system were utilized to monitor past forest fires. The Level 2 A product of Sentinel-2 imagery in GEE inherently includes atmospheric correction via the Sen2Cor processor and cloud and shadow coverage, ensuring that the input data is georeferenced and radiometrically corrected. To determine the extent of the burn and identify burned areas, the normalized burn ratio (NBR) was utilized. This index uses the near-infrared (NIR) and shortwave-infrared (SWIR) wavelengths according to Eq. 1. Healthy vegetation reflects strongly in the NIR but not in the SWIR. Areas devastated by fire have the opposite reflectance. A fire trail with burned vegetation and soil reflects strongly in the SWIR portion of the spectrum⁶⁸.

1

where NIR corresponds to Sentinel-2 band 8, and SWIR corresponds to band 11. NBR varies between − 1 and 1. High NBR values indicate healthy vegetation. Low values indicate barren land and recently burned areas. Unburned areas typically have values close to zero. To calculate this index, NBR is computed for images before and after the wildfire, and the post-fire image is subtracted from the pre-fire image to obtain the final value of this index. The dates of wildfires in Kurdistan Province from 2018 to 2022. Moreover, locations with burn scars were converted into 821 points for modeling, validating, and generating a wildfire susceptibility map (WSM) (Fig. 3b).

Landslide inventory map

For monitoring landslides and land subsidence, Sentinel-1 satellite images from 2015 to 2022 were employed using InSAR time series. The LiCSBAS (https://github.com/yumorishita/LiCSBAS) open-source software in Python was utilized for processing Sentinel-1 images and determining the interferograms. LiCSBAS uses LiCSAR products, eliminating the need for users to generate interferograms from single look complex (SLC) data. The generated interferograms use a gamma spectral filter for noise reduction. The LiCSAR tool utilizes the SNAPHU (Statistical-Cost, Network-Flow Algorithm for Phase Unwrapping) software (SNAPHU) for phase unwrapping. Coherent interferograms areas, where the filtered interferometric phase noise is less than 0.5, are masked during the phase recovery process. Encoded and converted to GeoTIFF format are the wrapped interferograms, phase recovery, and coherence images with a pixel spacing of 0.001 degrees (~ 100 m). The produced GeoTIFF images are downloadable and compatible with LiCSBAS software. The InSAR time series processor workflow is split into two parts. Phase 1 involves preparing batches of raw data, followed by Phase 2, which focuses on time series analysis.

Phase 1: Initially, LiCSBAS LiCSAR products for the study area are downloaded, and data format conversion is performed. Unwrapped phase GeoTIFF files from the COMET-LiCS system, covering a predefined geographical area (usually 250 × 250 km) and including both ascending and descending data for each time interval, are obtained. For enhanced accuracy and effectiveness, tropospheric noise correction is implemented on InSAR data using the general atmospheric correction online service (GACOS), followed by unwrapping the data.

Phase 2: Time series analysis uses the coverage and coherence of unwrapped data to detect and delete poorly unwrapped data that could affect the results. We invert purified unwrapped data to get displacement time series and velocity estimates. Then, we determine the standard deviation of velocity. Noisy pixels are then removed. Loop phases are computed from three interferometric sets, and data with significant errors are identified and removed from the processing. The loop phase is calculated using Eq. 2.

2

The loop phase should be approximately zero if there is negligible retrieval error in each of the three interferograms. Subsequently, the baseline shifts are inverted for the small baseline network to obtain displacement and velocity time series using the new small baseline subset (NSBAS) approach. The next step calculates the velocity standard deviation using the bootstrapping method. Subsequently, various noise indices obtained in the previous steps create a mask for the displacement time series and velocity. The pixel is removed if any noise index values for a pixel are greater/smaller than the specified threshold. The removal of various noises from images is shown in Fig. 4.

Fig. 4.

Indices of applied noise in the study area: Average coherence (Coh_avg), Number of data points used in unwrapped data (n_unw), Standard deviation of velocity (millimeters per year) (vstd), Maximum time length of the network (years) (maxTlen), Number of gaps in the network (n_gap), Spatial-temporal coherence (millimeters) (stc), Number of interferograms without loops (n_ifg_noloop), Number of unclosed loops (n_loop_err), Remaining error in baseline inversion (millimeters) (resid), and Velocity (vel).

To extract the filtered time series and velocity, the time series is subjected to a spatiotemporal filter (time-domain low-pass and spatial-domain high-pass with a Gaussian kernel, similar to StaMPS) that reduces the remaining noise. After generating the displacement velocity map, the locations of landslide occurrence are determined using the combination of three maps: displacement velocity, past landslide occurrence potential, and global sensitivity index (GSI). GSI is a geometric feature based on radar and surface geometry that assesses the potential for landslide occurrence. It can be calculated without the computational burden of downloading and processing large amounts of radar data, making it computationally efficient and particularly useful in the early stages of landslide susceptibility planning⁶⁹. Additionally, for creating the potential map of past landslide occurrence, 283 past landslide occurrence points in Kurdistan province, along with four metrics (Altitude, slope, slope aspect, and normalized difference vegetation index (NDVI)), are used. A linear regression method is applied to produce this potential map. Finally, by combining the global sensitivity index map, the past landslide potential map, and positive displacement velocity values, landslide occurrence locations in Kurdistan province are determined and identified. In the next step, to model and prepare a landslide susceptibility map (LSM), the landslide occurrence locations are transformed into points 1100 points for modeling and validation (Fig. 3c).

Land subsidence inventory map

To monitor land subsidence, Sentinel-1 satellite images from 2015 to 2022 are used. InSAR analysis is employed as a time-series method. The land subsidence monitoring steps and related analyses are similar to the landslide monitoring described in the previous section. The displacement velocity values indicate a maximum land subsidence of 109 mm per year in Kurdistan province. For modeling and preparing a land subsidence susceptibility map (LSSM), land subsidence occurrence locations are transformed into 1143 points (Fig. 3d).

Effective factors

In this study, nineteen influential criteria were considered independently by examining past studies for modeling and preparing multi-hazard susceptibility maps^{39,53,70–72}. These criteria include altitude, slope, slope aspect, distance from roads, rainfall, distance from villages, temperature, wind effect, distance from rivers, drainage density, distance from faults, land cover, topographic wetness index (TWI), soil type, groundwater drawdown, plan curvature, profile curvature and NDVI. Given that each criterion has a different impact on hazards, Table 1 outlines the effect of each criterion on risks.

Table 1.

Factors influencing each hazard.

Criteria	Land subsidence	Landslide	Flood	Wildfire
Altitude	✓	✓	✓	✓
Slope	✓	✓	✓	✓
Aspect	✓	✓	✓	✓
Distance to road	✓	✓	–	✓
Rainfall	✓	✓	✓	✓
Distance to village	–	–	–	✓
Temperature	–	–	–	✓
Wind effect	–	–	–	✓
Distance to river	✓	✓	✓	–
Drainage density	–	✓	✓	–
Distance to fault	✓	✓	–	–
Land cover	✓	✓	✓	✓
Topographic wetness index (TWI)	–	✓	✓	–
Geology	✓	✓	✓	–
Soil type	✓	✓	✓	–
Groundwater drawdown	✓	–	–	–
Plan curvature	✓	✓	✓	–
Profile curvature	✓	✓	✓	–
Normalized difference vegetation index (NDVI)	✓	✓	✓	✓

Topographic indices, including slope, altitude, aspect, plan curvature, TWI, and profile curvature, were extracted from the DEM derived from SRTM imagery using the GEE platform. Distance from the river, drainage density, and distance from the road were derived from 1:50,000 scale topographic maps obtained from the National Cartographic Organization (NCO). Rainfall, temperature, and wind speed maps were generated from annual averages (2015–2022) collected by the Iranian Meteorological Organization (IMO) for eight synoptic stations in Kurdistan province. Subsequently, raster maps of these parameters were prepared using the Kriging interpolation method. A wind effect index combining altitude and wind speed was utilized with the Windward/Leeward index to account for altitude changes in wind speed. A raster map depicting distances from villages was derived from a village distribution map obtained from the Iranian National Geosciences Database and converted to raster format using Euclidean distance. Geological map and distance from fault data were extracted from Geological maps of Iran at a scale of 1:100,000. Land cover and NDVI maps were derived from Landsat-8 satellite imagery from 2022 using the GEE platform. Soil-type data were obtained from Iran’s Natural Resources and Watershed Management Organization at a scale of 1:100,000 and converted to raster for modeling. Groundwater drawdown data were obtained from the water level of wells in the study area provided by the Iran Water Resources Management Company and interpolated using the Kriging method. All analyses and preparation of indices were conducted in ArcGIS 10.8 (https://www.esri.com) and SAGA GIS 8.2.1 (https://saga-gis.sourceforge.io/) software, with all indices considered at a pixel size of 30 × 30 m for modeling purposes (Fig. 5).

Fig. 5.

Factors influencing multi-hazards: (a) Soil type, (b) Altitude, (c) Slope, (d) NDVI, (e) Distance to fault, (f) Distance to road, (g) Rainfall, (h) Groundwater drawdown, (i) Distance to river, (j) TWI, (k) Plan curvature, (l) Profile curvature, (m) Land cover, (n) Aspect, (o) Geology, (p) Distance to village, (q) Temperature, (r) Wind effect, and (s) Drainage density (Maps are made using ArcGis10.8 (version: 10.8; URL link: https://www.esri.com)).

Preprocessing methods

Multicollinearity test

The association between independent variables in modeling can be determined by multicollinearity analysis⁷³. Two or more variables can be considered collinear if they depend significantly on one another. For this purpose, the variance inflation factor (VIF) index is used for multivariate analysis and is calculated through Eq. (3)²⁷.

3

where Tolerance is the tolerance level, and R² is the square of the R-squared of the regression. When the VIF is below 10, all parameters can be included in the model, and multicollinearity analysis is considered appropriate⁷⁴.

Frequency ratio (FR) method

The FR method is a bivariate statistical technique, a straightforward geographical tool for identifying possible relationships between independent and dependent variables⁷⁵. According to Razavi-Termeh et al.⁵⁸the FR technique considers the ratio of each hazard’s occurrences to the total occurrences of that hazard and the ratio of each class’s pixels to the total pixels in the study region (Eq. 4).

4

In this context, A represents the hazard-containing pixels per criterion, B represents the total occurrences of hazards in the study area, C represents the pixel counts within each criterion class, and D represents the overall pixel counts inside the study area.

Feature importance using regression relief method

A preprocessing technique used in machine learning, feature selection removes unnecessary and redundant features to enhance learning accuracy⁷⁶. The Relief method was formulated by Kira and Rendell in 1992, inspired by instance-based learning⁷⁷. Here are four steps to finding property weights using the Relief method: (1) Start weightless. (2) Randomly sample the training set. (3) Use Euclidean distance to find close hits and misses for each sample. (4) Calculate weight estimate. A sample with the closest neighbor from all samples of the same class is called a near hit, while a sample with the closest neighbor from all samples of different classes is called a near miss^78,79.

Modeling techniques

After data preparation, the dataset was split in a 70:30 ratio into training and testing sets using the Holdout method. Following preprocessing in MedCalc 23.2.0 software (https://www.medcalc.org/calc/), the data were further processed for modeling using the Python programming language in the Google Colab environment (https://colab.research.google.com/). Various libraries were used in this environment to model the RF model and optimize its five hyperparameters using PSO and GA algorithms. After modeling and developing the models for each hazard, the modeling prediction results were evaluated using standard performance metrics. Upon validation, susceptible areas were identified for all four hazards. The predicted data were exported from the Google Colab environment to ArcGIS 10.8, and multi-hazard susceptibility maps were prepared. Finally, MedCalc software evaluated the multi-hazard susceptibility maps based on various indicators.

Random forest (RF) algorithm

In 2001, Breiman developed the RF algorithm, a supervised learning method⁸⁰. In this algorithm, a large number of decision trees are formed, and decisions are made based on the majority vote as the final result of the algorithm⁸¹. This approach uses the Bagging process to make the findings more generalizable. Using the replacement of each chosen pattern, this method generates a new set of patterns for each tree iteration⁸². The sum of all the currently available patterns will determine the size of this sample collection. Approximately one-third of the preexisting patterns are usually removed by this sampling procedure. According to Lin et al.⁸³each tree can grow to a certain depth according to the pattern categorization. At each growth stage, variables are randomly selected, and the best branches are chosen.

Hyperparameter tuning using metaheuristic algorithms

This research utilized two metaheuristic algorithms, PSO and GA, to fine-tune the RF algorithm’s five hyperparameters. The GA and PSO algorithms, with various control parameters (Table 2), iterated multiple times to maximize the objective function, which is the R² score (coefficient of determination) (Eq. 5)³⁹.

5

Table 2.

Input parameters of metaheuristic algorithms in this study.

GA	PSO
Cross-over probability = 0.95 Mutation probability = 0.025 Population size = 50 Selection = Tournament Crossover = Uniform Mutation = Swap	Local coefficient (C1) = 2.05 Global coefficient (C2) = 2.05 Weight min of bird (W) = 0.4 Population size = 50

SS_res represent the squared differences between the actual observed values and the predicted values. SS_tot describes the squared differences between the actual observed values and the mean of the observed values. In conjunction with the RF algorithm, these two algorithms determine the optimal hyperparameters across different iterations and predictions tailored to the input dataset.

The following briefly explains the PSO and GA algorithms.

Particle swarm optimization (PSO) algorithm

Kennedy and Eberhart presented the PSO algorithm in 1995⁸⁴. This algorithm is based on animal group behavior, like how fish or birds fly in a flock. One way to look at PSO is as a swarm intelligence algorithm incorporating semi-evolution⁸⁵. To find the optimal solution, the algorithm randomly samples and evaluates them. Each chosen answer is linked to a search procedure in this algorithm that remembers its best positions up to this point⁸⁶. PSO uses the Eq. 6 to update the velocity for the optimization process⁸⁷.

6

where is the best personal memory of particle i at iteration t, is the best global memory of particle i at iteration t, W is the inertia factor, c₁ and c₂ are the personal and global learning coefficients, and r₁ and r₂ are random values between 0 and 1. The new position of particles in this algorithm is calculated based on Eq. (7)⁸⁷.

7

The particles’ new positions are determined by taking their prior positions and the updated velocities into account, after which the velocities are updated for each repetition.

Genetic algorithm (GA)

John Holland initially proposed the GA, a metaheuristic algorithm grounded in natural evolution principles^88,90. The GA mimics natural selection by allowing only the most vital individuals to persist. Natural selection favors more adapted individuals, who pass their genes along through the generations⁹⁰. Over time, this results in the prevalence of beneficial genes that help organisms better adapt to their environment in subsequent generations⁹¹. The GA starts by creating a starting population. Then, it uses an objective function to assess the fitness of each member of the population about a given problem⁹². A poorly defined objective function could lead the search astray and impact the ultimately produced solution, making its formulation very important. The fitness values are used to pick a subset of the chromosomes that will be the “parents” of the offspring. To create a new generation of humans, the genes of both parents are mixed and traded during mating⁹³.

According to Han and Xiao⁹⁴the algorithm believes that the offspring would inherit their parents’ best traits and outperform them. The GA consists of three processes: the first process is the selection of individuals to produce the next generation, and the subsequent processes involve manipulating the selected individuals to form the next generation using crossover and mutation techniques⁹⁵.

Validation techniques

To examine the error between target data and data predicted by the RF-hybrid algorithm, the root mean square error (RMSE) method was employed⁹⁶. This metric reflects the difference in the average of squared errors between actual and predicted data, providing a quantitative assessment of the model’s performance⁹⁷ (Eq. 8).

8

where n represents the number of data points, p_i denotes the predicted value by the model, and o_i represents the observed value.

This work evaluated the final map of areas susceptible to multi-hazards using the receiver operating characteristics (ROC) curve and the area under the curve (AUC). According to Marjanović et al.⁹⁸the ROC curve illustrates the compromise between the true positive rate (TP) and the false positive rate (FP). Moreover, the AUC, ranging between 0 and 1, indicates the accuracy and overall performance of the model (Eq. 9)⁹⁹.

9

where, P and N represent the total number of pixels associated with the occurrence and non-occurrence of the hazard, respectively.

Results

Result of multicollinearity test

Table 3 summarizes information about multicollinearity for each of the four hazards. The highest VIF index for flood is associated with the rainfall criterion (2.11), while the lowest is associated with drainage density (1.02). In the case of wildfire hazard, the highest VIF is related to the rainfall criterion (1.16), and the lowest is associated with the aspect criterion (1.005). The rainfall criteria and distance to roads have the highest and lowest values of the VIF for landslide, with values of 3.07 and 1.04, respectively. For land subsidence hazard, the highest VIF is associated with the rainfall criterion (2.39), while the lowest VIF is related to the aspect criterion (1.02). According to the results, the VIF value for all four hazards is less than 10, indicating that all criteria can contribute to modeling and susceptibility mapping.

Table 3.

Results of multicollinearity analysis on factors affecting multi-hazards.

Criterion	Land subsidence	Landslide	Wildfire	Flood
Altitude	1.31	1.46	1.11	1.16
Aspect	1.02	1.17	1.005	1.03
Drainage density	–	1.08	–	1.02
Geology	1.7	1.42	–	1.16
Land cover	1.36	1.67	1.04	1.19
NDVI	1.14	1.74	1.15	1.17
Plan curvature	1.33	1.27	–	1.15
Profile curvature	1.3	1.25	–	1.16
Rainfall	2.39	3.07	1.16	2.11
Slope	2.02	1.94	1.04	1.04
Soil type	1.8	2.07	–	1.66
TWI	–	2.52	–	1.58
Distance to rivers	1.04	1.06	–	–
Distance to villages	–	–	1.01	–
Wind effect	–	–	1.02	–
Temperature	–	–	1.09	–
Distance to roads	1.03	1.04	1.008	–
Distance to faults	1.23	1.17	–	–
Groundwater drawdowns	1.59	–	–	–

Assessing criteria class impact on hazard occurrences

Table 4 shows the results of the FR method’s evaluation of the spatial relationship between places of hazard occurrence and influencing criteria. This method assigns weights to different criteria classes, with higher weights indicating greater significance in modeling.

Table 4.

Results of frequency ratio method on factors affecting four hazards.

Criteria	Class	Flood	Wildfire	Landslide	Land subsidence
Altitude (m)	< 1545	1.6	1.5	2.96	0.86
	1545–1771	1.12	0.87	1.44	1.17
	1771–1977	0.58	0.76	0.61	1.73
	1977–2241	1.37	1.01	0.4	0.13
	> 2241	0.37	1.72	0.31	0
Slope (degree)	< 6	1.9	1.06	0.02	2.06
	6–13	0.99	0.97	0.28	0.62
	13–21	0.16	0.84	1.18	0.24
	21–30	0.05	1.14	2.67	0.17
	> 30	0.75	0.87	6.8	0.3
Aspect	Flat	0	2.17	1.6	0
	North	0.93	1.16	0.98	1.22
	North East	1.08	0.74	2.63	1.26
	East	0.87	1.15	2.19	1.35
	South East	1.4	0.93	1.12	1.08
	South	0.91	0.92	0.14	0.75
	South West	0.83	0.89	0.22	0.47
	West	0.99	0.99	0.08	0.78
	North West	0.92	1.17	0.16	1.02
Distance to roads (m)	0–200	–	1.39	1.07	1.24
	200–400	–	0.84	0.56	1.15
	400–600	–	1.03	0.57	0.9
	600–800	–	0.96	0.43	1.18
	> 800	–	0.98	1.06	0.97
Rainfall (mm)	231.9–405.2	1.58	0.62	0.003	1.88
	405.2–501.3	1.22	0.91	0.16	0.79
	501.3–620.4	0.16	1.43	1.76	0.068
	620.4–757.8	0.14	1.83	2.77	0.405
	> 757.8	0.54	0.74	4.76	0
Temperature (°C)	8.4–10.2	–	1.08	–	–
	10.2–11.2	–	0.8	–	–
	11.2–12.07	–	1.19	–	–
	12.07–12.82	–	1.01	–	–
	> 12.82	–	0.88	–	–
Wind effect	0.74–0.92	–	0.56	–	–
	0.92–1.01	–	1.03	–	–
	1.01–1.1	–	1.006	–	–
	1.1–1.2	–	1.13	–	–
	> 1.2	–	1.15	–	–
Distance to rivers (m)	< 100	1.09	–	0.92	1.07
	100–200	0.96	–	1.47	1.12
	200–300	0.97	–	1.21	1.05
	300–400	0.97	–	1.03	0.96
	> 400	0.99	–	0.81	0.93
Drainage density	0.02–0.23	0.53	–	0.43	–
	0.23–0.32	0.93	–	0.72	–
	0.32–0.4	1.07	–	1.26	–
	0.4–0.49	1.02	–	1.26	–
	> 0.49	1.38	–	0.87	–
Distance to faults (m)	0–300	–	–	2.23	0.18
	300–600	–	–	1.97	0.19
	600–900	–	–	1.74	0.28
	900–1200	–	–	2.09	0.3
	> 1200	–	–	0.75	1.18
Land cover	Range land	0.6	0.71	0.62	0.62
	Water area	0	1.01	1.51	0
	Farm land	1.4	1.18	0.41	1.4
	Forest	0.13	1.11	6.63	0.12
	Urban areas	0.73	0.58	0	0
	Uncovered plain	0	0	0	0
Topographic wetness index	− 12.5–2.5	0	–	0	–
	2.5–4.9	0.06	–	2.62	–
	4.9–6.2	0.57	–	1.008	–
	6.2–7.3	1.72	–	0.06	–
	> 7.3	1.95	–	0.007	–
Geology	Triassic	1.85	–	5.74	0
	Quaternary	0.81	–	0.05	2.1
	Precambrian	1.07	–	0.01	0.19
	Permian	1.09	–	0	2.49
	Paleocene	0.97	–	1.12	0.64
	Oligocene	1.24	–	2.25	0
	Oligocene	1.5	–	1.25	0.46
	Miocene	0.9	–	1.4	1.4
	Cretaceous	0.4	–	0	0.4
	Jurassic	0.73	–	1.33	0.29
	Eocene	0.82	–	0	0.41
	Paleozoic	1.02	–	0	0
	Cambrian	0	–	0	0
Soil type	Inceptisols	0.91	–	0.106	1.92
	Entisols/ Inceptisols	1.07	–	2.008	0.17
	Water body	0.58	–	0.08	0
	Rock outcrops	0.59	–	0	0.39
	Entisols/ Rock outcrops	1.12	–	0	1.97
Groundwater drawdowns (m)	0–6.5	–	–	–	0.58
	6.5–15.6	–	–	–	1.06
	15.6–1.8	–	–	–	2.47
	3.8–52.5	–	–	–	9.77
	> 52.5	–	–	–	10.82
Plan curvature	Convex	0.35	–	1.28	0.28
	Flat	1.29	–	0.91	1.28
	Concave	0.22	–	1.13	0.31
NDVI	− 0.59–0.08	3.72	1.6	1.32	0
	− 0.08–0.16	1.33	0.48	0.29	1.4
	0.16–0. 25	0.66	1.4	1.08	0.48
	0.25–0.41	0.55	1.5	3.31	1.09
	> 0.41	0.78	1.76	2.92	1.32
Profile curvature	− 0.01 – − 0.001	0.98	–	2.9	0.11
	− 0.001 – − 0.0005	0.94	–	1.43	0.43
	–0.0005–0.0003	0.95	–	0.53	1.44
	0.0003–0.001	1.21	–	1.23	0.69
	> 0.001	1.2	–	1.73	0.27
Distance to villages (m)	< 500	–	0.62	–	–
	500–1000	–	1.23	–	–
	1000–1500	–	0.91	–	–
	1500–2000	–	1.04	–	–
	> 2000	–	0.99	–	–

For floods, analysis of FR reveals that slopes less than 6 degrees exhibit the highest ratio (FR = 1.9) while southeast-facing aspects have the highest FR value (FR = 1.4). Additionally, higher FR is associated with TWI classes greater than 3.7 (FR = 1.95) and altitudes less than 1545 m (FR = 1.6). Notably, the 0–100 m distance from the river’s class has the highest value (FR = 1.09). In terms of rainfall, the class ranging from 231.9 to 450.2 mm holds the highest weight (FR = 1.58), while higher drainage density classes (> 49.0) have a FR of 1.38. The Triassic geology unit exerts the most influence on floods (FR = 1.85), and agricultural land cover is assigned the highest weight (FR = 1.4). Furthermore, the Entisols/ Rock outcrops soil classes significantly impact flood occurrence (FR = 1.12). In the NDVI criterion, the vegetation cover class ranges from − 0.59 to -0.08 and is weighted to 3.72. In the criteria of plan curvature and profile curvature, the classes flat (FR = 1.29) and 0.0003–0.001 (FR = 1.21), respectively, have the highest weights.

For wildfires, proximity to villages within the 500–1000 m range holds the highest weight (FR = 1.23). Rainfall levels between 620.4 and 757.8 mm carry the most significant influence (FR = 1.83), while agricultural land cover is predominant in land cover criteria (FR = 1.18). Higher Altitude (> 2241 m) have a significant impact on wildfires (FR = 1.72), and flat slope aspects are associated with increased occurrence (FR = 2.17). The steepest slope classes (21–30 degrees) carry the highest weight (FR = 1.14), while temperatures between 11.2 and 12.07 degrees Celsius hold significance (FR = 1.19). Wind effects are notable in classes more significant than 1.2 (FR = 1.15), and proximity to roads within the 0–200 m range is influential (FR = 1.572). Based on the NDVI results, the > 0.41 class is the most important (FR = 1.76).

Concerning landslides, slopes greater than 30 degrees exhibit the highest weight (FR = 6.8), with northeast aspect regions being the most affected (FR = 2.63). Lower altitudes (< 1545 m) have the highest weight (FR = 2.96), and lower TWI classes (2.5–4.9) hold significance effect (FR = 2.62). Additionally, higher NDVI classes (0.25–0.41) (FR = 3.31) and rainfall levels greater than 757.8 mm (FR = 4.76) are influential. Landslide occurrence is prominent within 0–200 m distances from rivers (FR = 1.47) and along distance between 0 and 200 m and 0–300 m from roads and faults (FR = 1.7 and FR = 2.23, respectively). The convex class in the plan curvature criterion has the greatest FR value of 1.28. The highest values for the geology and land cover factors are related to the Oligocene units and forest land zones, with FR values of 2.25 and 6.63, respectively. The highest importance in the drainage density criterion is attributed to the 0.32–0.49 class, with a FR value 1.26. The Entisols/ Inceptisols class in the soil criterion has the highest weight (FR = 2.008) and impact on landslide occurrence. The profile curvature criterion class with the highest FR value, at 2.9, is − 0.01 to − 0.001.

For land subsidence, altitudes between 1771 and 1977 m hold the highest association (FR = 1.73). Gentle slopes (< 6 degrees) and east-facing aspects are influential (FR = 2.06 and FR = 1.35, respectively). Flat plan curvature and greater groundwater drawdowns (> 52.5 m) are significant classes (FR = 1.28 and FR = 10.82, respectively). Proximity to rivers (< 100 m) and higher distanced to faults (> 1200 m) also play roles (FR = 1.12 and FR = 1.18, respectively). Entisols/ Rock outcrops soil classes and rainfall levels between 231.9 and 405.2 mm exhibit influence (FR = 1.97 and FR = 1.88, respectively). In contrast, the Permian geology formation and agriculture land cover classes hold significance (FR = 2.49 and FR = 1.4, respectively) in their respective parameters. Additionally, NDVI classes between − 0.08 and 0.16 are associated with higher occurrences of land subsidence (FR = 1.4). The highest values for the distance to road and profile curvature factors are related to the 0–200 m and − 0.0005 − 0.0003 classes, with FR values of 1.24 and 1.44, respectively.

Result of criteria importance

The results regarding the importance of criteria using the RRelief method for each of the four hazards are showed in Fig. 6. For the flood hazard, criteria such as altitude (0.049), NDVI (0.02), and drainage density (0.02) are of utmost importance. Rainfall (0.066), altitude (0.033), and temperature (0.029) are the most significant criteria for wildfire occurrence. Moreover, rainfall (0.203), land cover (0.186), and TWI (0.168) are significant in landslide. Similarly, for land subsidence phenomena, criteria like altitude (0.12), groundwater drawdowns (0.074), and NDVI (0.068) exhibit considerable significance. The results indicated that altitude, rainfall, and NDVI had a greater impact on the four hazards in the study area.

Fig. 6.

Importance of factors affecting multi-hazards.

Result of modeling

To model and prepare susceptibility maps for each of the four hazards, the RF algorithm was utilized and enhanced using two optimization algorithms: GA and PSO. As a result, the optimization algorithms were used to fine-tune five RF algorithm hyperparameters: maximum tree depth (max_depth), number of features to be considered for the best split (max_features), minimum number of samples needed to split an internal node in a tree (min_samples_split), minimum number of samples needed for a leaf node (min_samples_leaf), and the number of trees (n_estimators). Combined algorithms were modeled using the Python programming language in the Google Colab environment.

The susceptibility maps for the four hazards were prepared and implemented using the weights acquired from the FR approach as input for modeling. The R² score values for the GA and PSO algorithms in optimizing the RF algorithm were 0.91 and 0.87 for floods, 0.83 and 0.97 for wildfires, 0.96 and 0.952 for landslides, and 0.9 and 0.93 for land subsidence, respectively. The results indicate higher accuracy of the GA compared to the PSO algorithm in optimizing the RF algorithm for floods, wildfires, and landslides. Additionally, in optimizing the RF algorithm, the PSO algorithm demonstrated higher accuracy than the GA for land subsidence. The optimized hyperparameter values of the RF algorithm by the two optimization algorithms for the four hazards are presented in Table 5.

Table 5.

Optimized hyperparameters of random forest in modeling multi-hazards.

	Land subsidence	Landslide	Wildfire	Flood
GA	max_depth = 15	max_depth = 20	max_depth = 15	max_depth = 20
	max_features = 1	max_features = 5/0	max_features = 5/0	max_features = 5/0
	min_samples_split = 4	min_samples_split = 4	min_samples_split = 9	min_samples_split = 8
	min_samples_leaf = 8	min_samples_leaf = 1	min_samples_leaf = 8	min_samples_leaf = 1
	n_estimators = 30	n_estimators = 10	n_estimators = 30	n_estimators = 30
PSO	max_depth = 24	max_depth = 10	max_depth = 26	max_depth = 19
	max_features = 1	max_features = 6	max_features = 9	max_features = 5
	min_samples_split = 5	min_samples_split = 5	min_samples_split = 10	min_samples_split = 9
	min_samples_leaf = 2	min_samples_leaf = 1	min_samples_leaf = 10	min_samples_leaf = 1
	n_estimators = 26	n_estimators = 22	n_estimators = 17	n_estimators = 26

Table 6 presents the modeling results for the four hazards, using three algorithms: RF, RF-GA, and RF-PSO, based on the RMSE metric. Based on the results in the training dataset, the RMSE values for the RF, RF-PSO, and RF-GA algorithms for flood were 0.334, 0.304, and 0.315, respectively. For wildfires, these values were 0.453, 0.423, and 0.416; for landslides, 0.233, 0.224, and 0.220; and land subsidence, 0.318, 0.277, and 0.303, respectively. In the test dataset, the RMSE values for the RF, RF-PSO, and RF-GA algorithms for flood were 0.346, 0.333, and 0.331, respectively. For wildfires, these values were 0.478, 0.462, and 0.450, for landslides 0.238, 0.229, and 0.225, and for land subsidence 0.413, 0.362, and 0.344, respectively. The results indicated that the accuracy of the combined algorithms (RF-PSO and RF-GA) was higher than that of the standalone RF algorithm in all four modeling scenarios of hazards. Furthermore, the results indicated that modeling landslide, land subsidence, flood, and wildfire hazards yielded higher accuracy.

Table 6.

Results of modeling evaluation for four hazards.

Models	Land subsidence		Landslide		Wildfire		Flood
	Train	Test	Train	Test	Train	Test	Train	Test
RF	0.318	0.413	0.233	0.238	0.453	0.478	0.334	0.346
RF-PSO	0.277	0.362	0.224	0.229	0.423	0.462	0.304	0.333
RF-GA	0.303	0.344	0.220	0.225	0.416	0.450	0.315	0.321

Multi-hazard susceptibility maps

After developing combined models, they were applied to the full study area. Using ArcGIS 10.8 software, susceptibility maps for the four hazards were constructed. The susceptibility maps were categorized into five susceptibility classes: very low, low, moderate, high, and very high, using a natural break classification method. The susceptibility maps for the four hazards, generated by three algorithms, RF, RF-PSO, and RF-GA, are illustrated in Fig. 7a-c. Hazard susceptibility maps indicated that the eastern part of Kurdistan province was affected by flood and land subsidence. In contrast, from north to south, the western part was more susceptible to wildfires and landslides. According to the results, the central regions of Kurdistan Province were not vulnerable to any hazards.

Fig. 7.

Susceptibility map of hazards with RF-based models: (a) RF, (b) RF-PSO, and (c) RF-GA (Maps are made using ArcGis10.8 (version: 10.8; URL link: https://www.esri.com)).

The susceptibility map with higher accuracy was selected after preparing the susceptibility maps for each hazard. These maps were combined to create a composite multi-hazard risk map (Fig. 8). The results revealed that 16.72% of the study area was prone to flooding. Additionally, 10.88% of the study area was susceptible to wildfires, 10.51% to landslides, and 9.98% to land subsidence. Floods shared common areas with landslides, land subsidence, and wildfires in certain regions of Kurdistan province. Furthermore, wildfires overlapped with landslides in specific study areas. The combined analysis of floods, landslides, and wildfires demonstrated their co-occurrence in parts of Kurdistan province. In general, the overlap between land subsidence and flood was 5.07%, between landslides and wildfires was 2.12%, between flood and wildfires was 0.64%, between landslides and floods was 0.05%, and the overlap between floods, landslides, and wildfires was 0.015% of the study area. The primary hazard in Kurdistan province was flood, affecting numerous regions. Floods also exhibited significant overlaps with other hazards, particularly with land subsidence.

Fig. 8.

Susceptibility map of multi-hazards (Maps are made using ArcGis10.8 (version: 10.8; URL link: https://www.esri.com)).

The correlation between the four hazards in the study area is shown by the Pearson correlation coefficient in Table 7. Values close to 1 indicate a strong correlation between the two hazards, while values close to -1 indicate a very weak correlation between the two hazards. The results showed the highest correlation between flood and land subsidence, with a value of 0.69, followed by landslides and wildfires, with a value of 0.47. The correlation between other hazards was negative, indicating a weak correlation. It appeared that in Kurdistan province, flood and land subsidence, as well as wildfires and landslides, had a mutual influence on each other’s occurrence.

Table 7.

Correlation between high-risk areas in different hazards.

Hazard	Wildfire	Landslide	Land subsidence	Flood
Flood	− 0.43	− 0.63	0.69	1
Land subsidence	− 0.51	− 0.6	1	0.69
Landslide	0.47	1	− 0.6	− 0.63
Wildfire	1	0.47	− 0.51	− 0.43

Validation of multi-hazard susceptibility maps

To evaluate the susceptibility maps of hazards generated using three algorithms, namely RF, RF-PSO, and RF-GA, 30% of the occurrence data for each hazard and the ROC-AUC index were utilized (Fig. 9; Table 8). The results indicated that the AUC values obtained using the RF, RF-PSO, and RF-GA algorithms for flood were 0.89, 0.90, and 0.911, respectively. For land subsidence, the AUC values were 0.934, 0.959, and 0.951, respectively. Regarding wildfires, the AUC values were 0.838, 0.824, and 0.802, respectively. For landslides, the AUC values were 0.981, 0.989, and 0.991, respectively.

Fig. 9.

Validation using ROC curve in three models for hazards: (a) flood, (b) land subsidence, (c) landslide, and (d) Wildfire.

Table 8.

AUC results for four hazards in three models.

Hazard	Models	AUC
Flood	RF-GA	0.911
	RF-PSO	0.900
	RF	0.892
Land subsidence	RF-GA	0.951
	RF-PSO	0.959
	RF	0.934
Landslide	RF-GA	0.991
	RF-PSO	0.988
	RF	0.981
Wildfire	RF-GA	0.838
	RF-PSO	0.824
	RF	0.802

The ROC curve results demonstrated that the combined algorithms (RF-PSO and RF-GA) outperformed the RF algorithm regarding accuracy for all four hazards. Additionally, among the optimization algorithms, the GA performed better for floods, landslides, and wildfires. In contrast, the PSO algorithm performed better for land subsidence in optimizing the RF algorithm in modeling hazards.

Table 9 compares model results for different hazards, indicating significant differences in model performance. Across various hazard pairs, statistical tests revealed significant differences (p < 0.05) in AUC values, showing variations in model effectiveness. For instance, in comparing RF-GA and RF-PSO for flood hazard, the z-statistic was 3.22, suggesting a significant difference in AUC values (p = 0.0013). Similar significant differences were observed in comparisons between RF-GA and RF, as well as RF-PSO and RF for flood hazard. In summary, the results highlight variations in model performance across different hazards and underscore the importance of selecting appropriate modeling techniques for accurate susceptibility mapping.

Table 9.

Pairwise comparison of model results for different hazards.

Hazard	Models	Difference between areas	Standard error	z statistic	Significance level
Flood	RF-GA ~ RF-PSO	0.0108	0.003	3.22	P = 0.0013
	RF-GA ~ RF	0.0193	0.008	2.3	P = 0.0212
	RF-PSO ~ RF	0.0084	0.007	1.11	P = 0.2671
Land subsidence	RF-GA ~ RF-PSO	0.0079	0.002	3.35	P = 0.0008
	RF-GA ~ RF	0.0164	0.002	5.5	P < 0.0001
	RF-PSO ~ RF	0.0243	0.004	5.8	P < 0.0001
Landslide	RF-GA ~ RF-PSO	0.0025	0.001	2.2	P = 0.0278
	RF-GA ~ RF	0.0097	0.002	3.6	P = 0.0002
	RF-PSO ~ RF	0.0071	0.002	3.02	P = 0.0025
Wildfire	RF-GA ~ RF-PSO	0.014	0.011	1.19	P = 0.2319
	RF-GA ~ RF	0.0351	0.011	3.1	P = 0.0017
	RF-PSO ~ RF	0.0211	0.006	3.3	P = 0.0008

Discussion

Assessment of effective factors

In the wildfire hazard, factors such as high altitude, moderate slope, proximity to villages and roads, agricultural land cover, high temperatures, rainfall, wind effects, and flat slope aspect had a more significant impact on the occurrence of this hazard. In the study area, most forested areas are situated at higher altitudes, with higher elevations being more susceptible to wildfires. A decrease in the likelihood of wildfires with an increase in slope degree is attributed to rocky surfaces and the absence of fuel sources (i.e., vegetation) in this area¹⁰⁰. Near rural areas and roads, individuals have greater access to forests. Human activities in these areas, such as igniting fires for cooking, negligence in extinguishing them, and grass burning by farmers, can significantly influence the occurrence of wildfires¹⁰¹. Many of these fires in agricultural areas are intentionally set to clear crop residues; however, they should be managed to prevent their spread to adjacent forests and grasslands¹⁰². In the study area, altitude and rainfall were more critical than other factors in wildfire occurrence. The distribution of vegetation cover, fuel moisture content, rainfall, solar radiation, temperature, evaporation, wind speed and direction, and overall landscape features are all controlled by topography. These factors collectively affect the frequency and severity of forest fires¹⁰³.

The results of landslide hazard assessment indicated that steep slope, low altitude, northeast-facing aspect, high rainfall, proximity to rivers and roads, a distance of 1000 m from faults, forest land cover, dense vegetation, high drainage density, Triassic lithology unit, and Entisols/ Inceptisols soil had the most significant impact on landslide occurrence. In low slopes, instability does not occur extensively as resistant forces like soil friction usually surpass driving forces like gravity¹⁰⁴. However, in high slopes, climatic characteristics and vegetation cover create soil conditions conducive to instability at this degree of slope, especially when combined with increased gravitational force and human interventions⁹⁶. River erosion and bank abrasion disrupt slope equilibrium, leading to instability in riverbank slopes¹⁰⁵. Investigating rainfall’s role in landslide occurrences showed the highest correlation between rainfall levels exceeding 700 mm, with a strong relationship between rainfall amount and altitude. As the basin’s altitude increases, landslide occurrences increase, with the highest frequency ratio observed at an altitude of 1500 m. This trend underscores the role of rainfall in landslide occurrences, as increased altitude correlates with higher rainfall levels and increased landslide susceptibility¹⁰⁶. However, in high-altitude regions, the predominance of snowfall and frequent freezing phenomena during much of the year slow down soil formation process¹⁰⁷.

Additionally, the lack of sufficient water for saturation and reduced human interventions in nature at higher altitudes contribute to decreased landslide occurrences¹⁰⁸. Most landslides occurred in forested areas, attributed to more significant water infiltration and disregard for proper road construction principles in forested regions²⁵. Roads disrupt the natural state and, in effect, upset the area’s balance, leading to vertical incisions and increased pressure on the lower part of the road, consequently enhancing landslide occurrences near roads¹⁰⁵. Faults are tectonic fractures that typically alleviate pressure on rocks. However, their displacement and increased activities have significant effects, with landslides being one of the consequences near fault zones²⁵. According to the findings, the two most important factors in the study area’s landslide incidences were rainfall and land cover.

The results of land subsidence hazard assessment indicated that moderate altitude, low slope, east-facing aspect, low rainfall, agricultural land cover, low vegetation cover, proximity to rivers, high groundwater drawdowns, Permian lithological group, Entisols/ Rock outcrops soil, farther distances from faults, closer distances to roads, and smooth topography had a more significant impact on land subsidence occurrence in the study area. Faults trigger plate movements and subsidence in alluvial formations¹⁰⁹. The slope of the land significantly affects infiltration and, consequently, surface flow velocity¹¹⁰. Due to excessive water extraction, agricultural lands contribute to subsidence¹¹¹. Rainfall reduces land subsidence as heavy rainfall increases groundwater infiltration, raising groundwater levels¹¹². Depending on the neighboring surface effect, the slope aspect influences groundwater recharge¹¹³. Proximity to roads exacerbates subsidence by creating waves on the sides and soil disturbance¹⁰⁹. Proximity to rivers intensifies subsidence due to their impact on surface and groundwater¹¹⁴.

Furthermore, the results showed that altitude and groundwater drawdowns had a more significant influence on the occurrence of land subsidence than other factors. Altitude is crucial in land subsidence because it affects climatic factors like rainfall and groundwater. Overexploitation of groundwater resources leads to soil compaction, reduced voids, and land subsidence¹¹².

The results of flood hazard assessment revealed that in areas with low slope and altitude, southeast-facing aspect, high TWI, proximity to rivers, low rainfall, high drainage density, Triassic geological unit, agricultural land cover, Entisols/ Rock outcrops soil, sparse vegetation cover, and flat curvature, the likelihood of floods was higher. Because water flows from higher to lower locations in mountainous settings, watersheds at lower elevations are more likely to flood¹¹⁵. As stated by Al-Juaidi et al.¹¹⁶the majority of flooding occurs in riverbeds and then spreads to neighboring regions. The results show that the chance of flooding increases as the TWI rises. This is probably because lower infiltration rates lead runoff to be higher in places with higher TWI values¹⁹. Based on the plan curvature criterion results, the highest number of floods occurred in smooth terrain, likely due to its lower elevation position, which likely facilitates the acceptance and collection of flows and runoff¹¹⁵. The results indicated that altitude and NDVI significantly impacted flood occurrence in the study area.

Assessment of modeling

This study employed an optimization approach by combining RF machine learning models with GA and PSO algorithms to model and generate multi-hazard susceptibility maps. The results demonstrated good accuracy in modeling each hazard using this combined algorithm. The accuracy of the machine learning algorithms in modeling and producing hazard susceptibility maps was also notable. Machine learning algorithms operate without making assumptions in modeling and processing data based on their nature. In cases where data are incomplete or contradictory, machine learning algorithms perform well by recovering lost data based on existing patterns¹¹⁷. Machine learning algorithms are utilized to discover complex and nonlinear relationships among many parameters in modeling¹¹⁸.

This research utilized the RF machine learning algorithm, which exhibited good accuracy in modeling and generating multi-hazard susceptibility maps. The advantages of the RF algorithm are preventing overfitting, scalability, determining the importance of prediction parameters, and precise performance without extensive data¹¹⁹. The GA algorithm outperformed the PSO algorithm in accuracy among the metaheuristic algorithms. GA offer advantages over PSO in their ability to explore diverse solutions, robustness to changes in the problem space, adaptability to different problem types, parallelization for faster convergence, maintenance of diversity in populations, and handling of constraints more effectively¹²⁰.

Comparison with previous research

The accuracy of the susceptibility maps obtained in this study for flood, land subsidence, landslide, and wildfire hazards were 91.11%, 95.9%, 99.1%, and 83.8%, respectively. Compared to previous studies, Pourghasemi et al. (2020)¹²¹ reported accuracies of 83.4% for flood, 93.9% for landslide, and 94.3% for forest fire using the RF model. Pourghasemi et al. (2019)⁴² achieved 84% and 80% accuracy for flood and landslide hazards, respectively, using the ANFIS. Kim et al. (2025)¹²² reported 99%, 99%, and 95.8% accuracy for drought, flood, and forest fire hazards, respectively, using Extreme Gradient Boosting (XGB) and RF models. Using the RF model, Pourhashemi et al. (2025)¹²³ achieved accuracy of 93.5%, 91.4%, and 95.6% for dust, flood, and landslide hazards. Javidan et al. (2021)¹²⁴ reported accuracies of 93.6%, 88.5%, and 92% for flood, landslide, and gully erosion hazards, respectively, using the Maximum Entropy (MaxEnt) model. Bordbar et al. (2022)⁵² achieved 93.6% and 89.4% accuracy for flood and landslide hazards using the SWARA-ANFIS-PSO model. Rahmati et al. (2019)⁴⁹ obtained an accuracy of 92.4% for avalanche hazards using the SVM method, and accuracies of 93.7% for rockfalls and 94.2% for floods using the Boosted Regression Tree (BRT) method. Nachappa et al. (2020)¹⁹ reported 87% and 90% accuracy for flood and landslide hazards, respectively, using the RF model. Pourghasemi et al. (2023)⁵⁵ achieved 81%, 85%, and 94% accuracy for forest fire, flood, and landslide hazards, respectively. Overall, this study’s accuracy for flood, landslide, and land subsidence hazards is notably higher compared to most previous studies.

Causal relationships between hazards

In our research, the interaction between the hazards — floods, land subsidence, wildfires, and landslides — shows spatial overlap and correlations. The most significant positive correlation is between floods and land subsidence (0.69), likely due to similar environmental conditions, including groundwater depletion and shifts in land elevation. This may exacerbate land subsidence after floods, particularly in regions already impacted by groundwater extraction, where elevation changes in the recovery of the ground post-subsidence could further increase flooding risk and vulnerability¹²⁵. Land subsidence can similarly increase flooding risk by altering drainage patterns. In addition, a moderate positive correlation (0.47) between wildfires and landslides, suggesting that wildfires may contribute to landslide risk through vegetation degradation. After a wildfire, the loss of root structure reduces cohesion, further exposing soil to erosion in mountainous regions¹²⁶. Weakened soils increase the risk of landslides, particularly after significant rainfall events, in which increased water infiltration destabilizes the soil¹²⁷. In contrast, the correlation between floods and wildfires is weakly negative (-0.43), indicating these environmental hazards tend to occur in related ecological conditions. Floods generally occur more readily in regions with high rainfall or near river networks.

Future directions

For future research, several recommendations can be proposed. Firstly, it is suggested that LiDAR technology be utilized alongside radar and optical imagery. This combination of technologies has the potential to significantly enhance the accuracy of identifying and predicting various natural hazards, such as floods, landslides, and wildfires. By integrating LiDAR data with other remote sensing techniques, researchers can obtain more detailed and comprehensive information about terrain features and surface characteristics, leading to more precise hazard assessments. Secondly, incorporating geophysical data into hazard modeling is advised. Geophysical data, including information about subsurface structures, fault lines, and ground stability, can provide valuable insights into the underlying factors contributing to hazard occurrences. For example, hidden faults and land subsidence may have significant implications for specific hazards, and integrating such data can improve the understanding and prediction of these phenomena.

Furthermore, it is recommended that the application of additional machine learning and deep learning approaches in hazard prediction be explored. By leveraging advanced algorithms in combination with metaheuristic algorithms, researchers can achieve higher accuracy and reliability in predicting regional hazards. These methodologies can analyze complex datasets and identify non-linear relationships between various parameters, thereby enhancing the predictive capabilities of hazard models. Additionally, the integration of knowledge-based and data-driven methods is proposed. This hybrid approach can facilitate selecting and analyzing influential factors in the study area, considering expert knowledge and empirical evidence. By combining these approaches, researchers can develop more robust hazard susceptibility maps and decision support systems accessible to users through web and mobile platforms. This would enable stakeholders and decision-makers to access critical information rapidly and effectively, improving disaster preparedness and response efforts.

Conclusion

Due to its geographical location, Kurdistan Province is consistently exposed to various natural hazards due to its topographical, geological, and climatic features. Therefore, a combination of the RF algorithm with GA and PSO algorithms was employed to produce a susceptibility map for each hazard. The results showed that the application of metaheuristic algorithms significantly enhanced the RF model’s predictive accuracy. Specifically, the AUC index values for the RF-GA, RF-PSO, and baseline RF models were 91.1%, 90.0%, and 89.2% for flood prediction; 95.1%, 95.9%, and 93.4% for land subsidence; 99.1%, 98.8%, and 98.1% for landslide; and 83.8%, 82.4%, and 80.2% for wildfire, respectively.

The results indicated that altitude and NDVI were the most significant and influential in flood hazards. In contrast, altitude and rainfall played crucial roles in wildfire hazards. For landslides, rainfall and land cover were identified as the most important factors, while altitude and groundwater drawdowns were the most influential in land subsidence hazards. Moreover, the findings suggested that the eastern and western regions of Kurdistan province are more prone to various risks, while the central areas are less vulnerable. The eastern regions are more affected by floods and land subsidence, whereas landslides and wildfires predominantly impact the western regions. Additionally, areas in the southwestern parts of Kurdistan province are susceptible to multi-hazards. Among the hazards in Kurdistan province, floods and land subsidence showed a high correlation (0.69) and could mutually influence one another’s occurrence. Similarly, wildfires and landslides exhibited shared characteristics and could affect each other’s incidence (0.47). Flood hazards also demonstrated the highest correlation with other hazards in Kurdistan province. Overall, the study contributes valuable knowledge for enhancing disaster preparedness and mitigation efforts in Kurdistan Province, ultimately striving towards building a more resilient and safer environment for its inhabitants.

Author contributions

Seyed Vahid Razavi-Termeh: Data curation, Conceptualization, Investigation, Methodology, Software, Formal analysis, Visualization, Writing - Original Draft. Abolghasem Sadeghi-Niaraki: Investigation, Methodology, Project administration, Supervision, Validation, Writing - Review & Editing. Farman Ali: Validation, Resources, Methodology, Writing - Review & Editing. Biswajeet Pradhan: Validation, Methodology, Writing - Review & Editing. Soo-Mi Choi: Investigation, Methodology, Supervision, Project administration, Funding acquisition, Writing - Review & Editing. All authors read and approved the final manuscript.

Data availability

The data that support the findings of this study are available on request from the corresponding author.

Code availability

The code that supports the findings of this study is available upon request from the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.