1 Introduction

Wildfires pose a significant threat to forest ecosystems, human lives, and infrastructure on a global scale (Carvalho et al. 2021; Nones et al. 2024). The frequency and intensity of wildfires have increased in recent years due to climate change, land-use practices, and population expansion (Eslamani et al. 2018; Javadinejad et al. 2019a, b). Forests cover one-third of the Earth’s land, providing crucial habitats for biological diversity and ecological services. Forest fires are natural occurrences that benefit ecology but can cause human health hazards, soil erosion, air pollution, biodiversity loss, and financial costs (Nafchi et al. 2022; Geraskina et al. 2022). The significance of wildfires in biodiversity hotspots is influenced by their unique ecosystem characteristics and the interaction between fire and other ecological processes (Kganyago and Shikwambana 2019; Gajendiran et al. 2023). Global forest fires are most common in the wildland-urban interface (WUI), where structures are vulnerable to destruction and significant economic losses (Li et al. 2022; Nunes et al. 2023). Africa’s annual wildfires are estimated to have risen by 70% and 45% of forest fires and woodlands, causing significant socioeconomic and environmental impacts (FAO 2020; Kala 2023). The rise in forest fires has exacerbated the challenge of balancing conflicting vegetation management goals, especially regarding ecosystem services and biodiversity conservation (Afuye et al. 2021a; Dhyani 2023; Mpanyaro et al. 2024). Understanding the intricate dynamics of wildfires in biodiversity hotspots and wildfire risk assessment methodologies is crucial for effective management and balancing ecological benefits and risks.

Anthropogenic climate change is predicted to cause significant global habitat changes for numerous species in the Mediterranean Basin (Aurelle et al. 2022). Based on extensive observational data and scientific modelling, the Intergovernmental Panel on Climate Change (IPCC) report identifies forest fires as a significant contributor to climate warming (Masson-Delmotte et al. 2021). This issue poses a considerable challenge for humanity. Hybrid artificial intelligence, including neural-fuzzy inference, Grey Relativity Analysis method, Bayesian networks, and GIS-based Multi-Criteria Decision Analytic Hierarchy Process (MCDA-AHP), has been utilised in studies to evaluate forest fire risk (Javadinejad et al. 2019a, b; Balocchi et al. 2022; Rihan et al. 2023). This study utilised the GIS-based AHP to evaluate wildfire risk (Nguyen et al. 2024), considering human, topographic, and climatic factors to address the gap in the current methodology and develop proactive measures for managing ecologically sensitive areas. For instance, Gigović et al. (2018) used AHP to model forest fire hazards across 17,000 hectares in Bosnia and Herzegovina, Southeast Europe, revealing that land use, socioeconomic, topography, and climate had the highest values among the different clusters. Gai et al. (2011) assessed forest fire risk assessment and mapping in Southern China, focusing on hazard, vulnerability, and response capacity. They identified suitable land use types based on elevation, slope, aspect, temperature, wind, and relative humidity. Jiang et al. (2021) analysed forest fire hazards in northern Gävleborg County, Sweden, using historical fire point data, categorising factors into natural and human-caused, revealing that most fires occurred in sites with the highest vegetation type weight.

In South Africa, a study shows that the dry summer climate and associated vegetative fuel load contribute to wildfire ignition and spread, creating a high-risk zone for biodiversity hotspots and wildfire patterns (Strydom and Savage 2016). A study by Forsyth and Maitre (2015) revealed that wildfire poses a significant risk to over half of the 3700 buildings in Plettenberg Bay in the Western Cape. Kraaij et al. (2018) highlighted the Knysna region’s fire control challenges due to high fire danger, drought, fynbos vegetation conversion, and alien tree invasion, increasing risks for residents. Consequently, wildfire risk assessment studies in South Africa’s biodiversity hotspots still lack fire elements such as historical fire occurrences and wildfire behaviour factors (Forsyth and Maitre 2015; Strydom and Savage 2016; Kraaij et al. 2018; Xulu et al. 2021). Few studies have effectively used MCDA-AHP to identify high-risk biodiversity hotspots and assess wildfire risk using the differenced Normalized Burn Ratio (dNBR) index for essential elements affecting wildfire behaviour (Güngöroğlu 2017; Tshering et al. 2020; Maric et al. 2021; Vujović and Nikolić, 2022). Modelling spatial patterns of wildfire occurrence in South Africa’s biodiversity hotspots can aid precise decision-making for wildfire management, land-use planning, and emergency response.

Wildfires significantly impact South Africa’s Mediterranean vegetation type, characterised by a unique combination of fire-adapted habitats and species (Le Maitre et al. 2014; Bradshaw et al. 2018). The Garden Route District (GRD) in the Western Cape Province experiences a Mediterranean climate with average temperatures ranging from 25 to 19 °C, with warm summers and cool, wet winters influenced by the Agulhas Current (Wilson 2012). The region’s unique terrain, climate, and ecological factors significantly influence various vegetation types (Frost 1999; Frost et al. 2018). These include Fynbos, renosterveld, thicket, forest, woodland, and coastal dune vegetation, all influenced by their specific characteristics. The GRD is characterised by fire-prone vegetation and agricultural land (Frost et al. 2018; Tshering et al. 2020; Maric et al. 2021; Vujovic and Nikolić 2022). This, therefore, necessitates modeling spatiotemporal wildfire risk patterns to enhance resilience, promote conservation, regulate development in high-risk areas, and encourage fire-wise practices. The Mediterranean dry summer climate and vegetative fuel load have led to significant economic losses due to wildfires affecting vegetation cover types and plant growth over the past few decades (Kruger et al. 2006; Omarjee 2022). This study aimed to model spatiotemporal patterns of wildfire risk in biodiversity hotspots from 1991 to 2021 using weight assignment and overlay analysis. Hence, the MCDA-AHP model was validated through statistical tests and Confusion Matrix. The outcome of this study can be used to identify fire-prone biodiversity environments, facilitate resource allocation, improve land management, and raise public awareness.

2 Material and methods

2.1 Study area

The Garden Route District (GRD) Municipality in the Western Cape Province, South Africa, is located between 33° 70′ 42′′ S and 22° 04′ 76′′ E (Fig. 1). The GRD has a maximum elevation of 555 m above sea level and covers 23,331 km2 with a population of 627,917 Fields (GRDM-IDP 2022). The microclimatic conditions in the district significantly impact the topography and ocean currents. The region has a Mediterranean climate with cool, wet winters and warm, dry summers influenced by the warm Agulhas Current (Wilson 2012). Summer temperatures range between 15 and 25 °C, while winter temperatures range between 8 and 19 °C (Xulu et al. 2021). The dominant vegetation types in the study area are the evergreen, hard-leaved Mediterranean shrubs, known as Fynbos, temperate forests, shrublands, and thicket vegetation (GRDM-IDP 2022). The fynbos vegetation is a biodiversity hotspot prone to fire, which can cause damage to surrounding areas (Topp et al. 2022). Fires typically occur from mid-October to April during the dry summer season, lasting about 23 weeks (Platt et al. 2015).

Fig. 1
figure 1

Map of Garden Route District Municipality in the Western Cape, South Africa, and the land cover types

Full size image

2.2 Data

The study utilised data from human, topographic, and climatic factors categorised as climatic risk components to develop the Analytic Hierarchy Process (AHP) model from 1991 to 2021. The study used eleven datasets from three categories to examine human influence, including population density, LULC, distance from settlements, aspect, slope, and road distance. The population density data was obtained from the Socioeconomic Data and Applications Centre (SEDAC) at 30 m resolution (SEDAC 2021). The LULC map was created from Landsat imagery at 30 m resolution and the South African Department of Forestry, Fisheries and the Environment (DFFE 2020). Distances from roads and settlements were extracted from the LULC map. The Digital Elevation Model (DEM) was obtained from OpenTopography at 30 m resolution (opentopo.com). The DEM was used to derive the slope and aspect, adding elevation as a topographic factor input into the AHP model. The DEM was used to calculate slope and aspect, incorporating elevation as a topographic factor into the AHP model, thereby enhancing the accuracy of the model predictions. WorldClim data was utilised to gather climate information, including precipitation, temperature, wind speed, and water vapour at 21 km and 1 km spatial resolutions (www.worldclim.org). The United States Geological Survey (USGS) provided Landsat data, which included Landsat 5 TM, Landsat 7 ETM+, and Landsat OLI/TRIS-2 (8–9) sensors at a 30 m resolution (https://earthexplorer.usgs.gov/). These data were used to map and compute the surface reflectance of the burnt area. In addition, historical fire records, in point format, were derived from MODIS data obtained from the Global Forest Watch (GFW) database (https://globalforestwatch.org/). The data was resampled to a 30 m resolution before integrating into the Analytic Hierarchy Process (AHP), as illustrated in Fig. 2.

Fig. 2
figure 2

Flow chart illustrating the wildfire risk assessment model

Full size image

2.3 Multi-criteria decision analysis-analytic hierarchy process (MCDA-AHP) model

This study employed the Analytic Hierarchy Process (AHP) to evaluate wildfire risk in biodiversity hotspots, involving data acquisition, pre-processing, thematic layer, weighted overlaying, fire risk map generation, and validation (Fig. 2). The data was analysed using spatial and temporal harmonisation, and thematic layers were created to classify each criterion into five risk levels based on the wildfire risk impact (Faramarzi et al. 2021). Previous studies have classified risk into low, low-to-moderate, moderate, moderate-to-high, and high-risk zones in the hierarchy of levels (Youcef et al. 2020; Faramarzi et al. 2021). Table 1 presents the risk levels that are ranked, with the highest risk being the highest (5) and the lowest risk being the lowest (1). The MCDA-AHP first step involved data acquisition, pre-processing, and the creation of thematic layers with risk classes (ranks). The second step involved assigning weights to all thematic layers using AHP, following the methods described by Satty (1987), Podvezko (2011), and Taherdoost (2018). The initial step involved creating a pairwise comparison matrix, comparing all factors (criterion) considered for the model based on their influence and importance in wildfire risk, as illustrated in Table 1.

Table 1 Pairwise comparison matrix
Full size table

The second step involved calculating the matrix product using Eq. (1).

$${\Pi }_{i}=\prod_{j=1}^{n}{p}_{ij}$$
(1)

where \({\Pi }_{i}\) is the product of each row in the pairwise matrix, \(i\) and \({p}_{ij}\) is the element of the pairwise comparison matrix, and n is the number of factors being compared (Podvezko 2011) in this case, n = 11. The third step involved computing the n-th degree root from the obtained products in step 2 using the formula in Eq. (2).

$$\sqrt[n]{{\Pi }_{i}}$$
(2)

The fourth step involved calculating the weights of each criterion by normalizing the obtained values from step three and dividing each criterion product by the sum obtained, as shown in Eq. (3).

$${\omega }_{i}=\frac{\sqrt[n]{{\Pi }_{i}}}{\sum_{i=1}^{n}\sqrt[n]{{\Pi }_{i}}}=\frac{\sqrt[n]{\prod_{j=1}^{n}{p}_{ij}}}{\sum_{i=j}^{n}\sqrt[n]{\prod_{j=1}^{n}{p}_{ij}}}$$
(3)

where \({\omega }_{i}\) is the criterion weight, \(i\) and \({p}_{ij}\) is the element of the pairwise comparison matrix (Podvezko 2011). The fifth and final step involved calculating the consistency of the pairwise matrix using Eqs. (4) and (5), with the resulting weights shown in Table 2.

$$CI=\frac{{\lambda }_{max}-n}{n-1}$$
(4)

where \(CI\) is the Consistency Index used to assess the consistency ratio. At the same time, \({\lambda }_{max}\) represents the maximum eigenvalue of the pairwise comparison matrix, and \(n\) represents the number of compared criteria. \({\lambda }_{max}\) is calculated by dividing the weights of each element by their weight and calculating the average of the resulting eigenvalues.

$$CR=\frac{CI}{RI}$$
(5)

where \(CR\) is the Consistency Ratio (CR), which is the ratio of the pairwise comparison matrix, while the Random Index (RI) is a function of n described by Satty (1987) AHP literature Table 2. For a matrix with n = 11, \(the RI=1.51\) and CR must be less than or equal to 0.1 for acceptance, indicating consistency (Podvezko 2011). Table 2 indicates that the CR for this study was equal to 0.1, indicating that the pairwise matrix was consistent.

Table 2 Weights and consistency ratio calculation
Full size table

Lastly, the weighted factors were combined to create a wildfire risk map using ArcGIS v10.8 software in the ArcGIS Environment. Figure 2 shows the study flowchart, while Table 3 provides a detailed list of processed thematic layers in the ArcGIS Environment, including their weights and ranks.

Table 3 Evaluation of wildfire risk based on the weight of each criterion
Full size table

2.4 Wildfire risk model validation

2.4.1 Validation with historical fires (points and burned area)

This study validated the wildfire risk assessment model using historical fire data from the Global Forest Watch database (https://globalforestwatch.org/), overlaying fire point occurrences from 1991 to 2021. The wildfire risk maps were categorised into five distinct classes: low, low-to-moderate, moderate, moderate-to-high, and high-risk zones. Historical overlapping fire points were counted in each risk zone to analyse the influence of each risk zone on fire occurrence. The model accuracy was assessed by comparing the burned area per risk zone to the mapped risk zones in the wildfire risk zonation map (Kraaij et al. 2018; Hao et al. 2022). Therefore, the burned area was mapped using Landsat data and the differenced Normalized Burn Ratio (dNBR), as computed in Eqs. (6) and (7).

$$NBR= \frac{NIR-SWIR}{NIR+SWIR}$$
(6)
$$dNBR= {NBR}_{pre}-{NBR}_{post}$$
(7)

where \(NBR\) is the Normalised Burn Ratio calculated before (\({NBR}_{pre}\)) and after (\({NBR}_{post}\)) the fire, while the NIR and SWIR are Near Infrared (NIR) and Shortwave Infrared (SWIR) bands used to calculate NBR (Hao et al. 2022). The \(dNBR\) is the actual dNBR calculated by subtracting the post-fire NBR from the pre-fire NBR and ultimately maps the burned scar using a threshold value that indicates the burned area (Kraaij et al., 2018). The threshold value for this study ranges between 0.10 and 0.12 but may vary depending on different locations (Franco et al. 2020). Consequently, the model underwent three adjustments to enhance its accuracy based on historical fire occurrences (Yang et al. 2021). The model was validated by overlaying all thematic layers with historical fire points based on wildfire risk, aligning with previous studies (Asori and Emmanuel 2020; Kayijamahe et al. 2020; Lamat et al. 2021; Yang et al. 2021). The objective was to monitor each thematic layer and its impact on wildfire occurrence and behaviour.

2.4.2 Validation with statistical tests

The model performance was validated through statistical tests, specifically the Confusion Matrix, a widely used tool for assessing wildfire risk accuracy (Yue et al. 2023). The analysis evaluated Confusion Matrix metrics such as Overall Accuracy, Sensitivity, Specificity, and Kappa Coefficient (Kappa’C). Malik et al. (2021) identified four key metrics, including True Positive (TP), True Negative (TN), False Positive (FP) and False Negative (FN). Overall Accuracy measures the proportion of correctly classified instances, while Sensitivity and Specificity measure the proportion of positive and negative risk categories correctly identified, indicating overall model performance (Malik et al. 2021; Yue et al. 2023). The Kappa’C measures the model consistency and randomness in predicting wildfire risk, with a higher value indicating greater model reliability (Sinha et al. 2023). The indicators (metrics) are calculated using the formulas in Eqs. (8) to (11).

$$Sensitivity= \frac{TP}{TP+FN}$$
(8)
$$Specificity= \frac{TN}{TN+FP}$$
(9)
$$Accuracy= \frac{TP+TN}{TP+FN+TN+FP}$$
(10)
$$Kapp{a}{\prime}C= \frac{Accuracy-{P}_{e}}{1-{P}_{e}}\left({P}_{e}=\frac{(TP+FN)(TP+FP)(TN+FN)(FP+TN}{{(TP+FN+FP+TN)}^{2}}\right)$$
(11)

where, TP stands for true positive, TN represents true negative, FP and FN represent false positives and false negatives, respectively, and Pe calculates the probability of the model making random predictions (Yue et al. 2023). The study utilised the Confusion Matrix to evaluate the effectiveness of the wildfire risk model in identifying wildfire risks in the study area. The risk rank was determined by the percentage of historical fires in the risk zones, similar to how influencing factors were reclassified, as shown in Table 4. The number of fires per risk zone ranged from 9% for low and low-moderate risks to 12% for moderate, 30% for moderate-high, and 40% for high-risk zones.

Table 4 Wildfire risk ranking with the number of fires per risk zone
Full size table

Furthermore, the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R2, with low MAE and RMSE values and high R2 values, were evaluated for further validation (Mestre-Runge et al. 2023). The model predictions are more accurate than actual wildfire occurrences by the values calculated using Eqs. (12) to (14).

$$MAE= \frac{1}{n}{\sum }_{i=1}^{n}\left|{y}_{i}-{\widehat{y}}_{i}\right|$$
(12)
$$RMSE=\sqrt{\frac{1}{n}{\sum }_{i=1}^{n}\left|{y}_{i}-{\widehat{y}}_{i}\right|}$$
(13)
$${R}^{2}=1-\frac{{SS}_{res}}{{SS}_{tot}}$$
(14)

where for both MAE and RMSE, n represents the number of observations, \({y}_{i}\) is the actual value, and \({\widehat{y}}_{i}\) represents the predicted value. In \({R}^{2}\), however, \({SS}_{res}\) represents the sum of squared residuals, and \({SS}_{tot}\) is the sum of squares (Mestre-Runge et al. 2023). This study used observations to represent the total number of wildfire risk zones, while actual values were historical fire points within each zone. The predicted value reflected the model’s forecasted count of fires for the same zones based on wildfire risk predictions.

2.5 Temporal patterns analysis

The study assessed temporal patterns of wildfire risk and burned area in the Garden Route, a biodiversity hotspot area, using time series, correlation analysis, and regression modelling. The study utilised a time series analysis to examine the temporal trends in burned areas and historical fire occurrences from 1991 to 2021. The correlation between burned area, fire seasons, and historical fire occurrences was evaluated using a correlation matrix, with − 1 indicating a perfect negative correlation and 1 indicating a perfect positive correlation (Busico et al. 2019). The regression analysis involved fitting a model with historical fire occurrences as the response variable. The burned area and occurrences are predictors, with a p-value of 0.05 or lower indicating statistical significance (Cizungu et al. 2021). This study uses burned areas or fire seasons as predictors, with higher R2 values, indicating that the regression model effectively explains historical fire occurrences (Parajuli et al. 2020).

3 Results

3.1 Thematic maps overlaid with Historical fire occurrences

The historical fire points were overlayed on all the thematic maps to enhance the overall validation of the model (Fig. 3). Historical fire occurrences revealed the population density (Fig. 3a), distance from settlements (Fig. 3b), LULC (Fig. 3c), and distance from roads (Fig. 3d). The results show numerous fire points near areas with high population density, as shown in Fig. 3a. The distance from settlements and the distance from roads (Figs. 3b and 3d) show concentrated fire points in shorter distances (red) except for the northern regions of the study area. The LULC map shows that the fire points are concentrated in vegetated and cultivated regions close to the built-up areas, where the LULC types represent high risks.

Fig. 3
figure 3

Historical fire occurrences overlaid on human influence thematic maps

Full size image

Figure 4 shows the correlation between fire occurrences and topographic maps, including the aspect (Fig. 4a), elevation (Fig. 4b), and slope (Fig. 4c). Aspect in Fig. 4a shows that fire points are mainly located on North and West facing slopes, with a notable concentration of fire points on South facing slopes. Elevation in Fig. 4b indicates that most fire points are concentrated in low-lying areas, marked by the red colour, indicating a high-risk zone. The slope (Fig. 4c) thematic layer reveals that fire points are mainly concentrated in moderate-to-high-risk and high-risk zones at uphill angles between 31°–46°, 46°–62°, and 62°–77°, respectively.

Fig. 4
figure 4

Historical fire occurrences overlaid on topographic influence thematic maps

Full size image

Figure 5 shows the correlations between climatic factors and wildfire occurrences, focusing on precipitation (Fig. 5a), temperature (Fig. 5b), water vapour (Fig. 5c), and wind speed (Fig. 5d). Precipitation indicates that higher fire point areas were found in low rainfall areas, indicating a high-risk zone (Fig. 5a). The study reveals that areas with higher temperatures had more fire points, indicating moderate-to-high-risk areas (Fig. 5b). Nevertheless, the moderate risk zones are identified by areas with mild temperatures between 16 and 18 °C that exhibit noticeable fire points. Water vapour exhibited the highest fire points in moderate-to-high-risk and high-risk zones between 1.4–1.6 kPa and 1.6–1.8 kPa, as depicted in Fig. 5c. The wind speed in Fig. 5d indicates that fire points are primarily located in mid-wind speed (3.3–3.9 m/s) and high-risk zones (4.5–5.0 m/s), with a significant portion in high-wind speed areas.

Fig. 5
figure 5

Historical fire occurrences overlaid on climatic influence thematic maps

Full size image

3.2 Wildfire risk

The wildfire risk zonation map was created using MCDA-AHP and historical fire occurrences based on the weighted overlay of fire influencing factors, as shown in Fig. 6. Figure 6a shows a wildfire risk zonation map, while Fig. 6b shows the percentage distribution of risk zones within the study area, displaying five fire risk levels. However, the study area is primarily occupied by the low-moderate-risk zone, with 27% in the northwestern parts and 26% in the high-risk zone in the southwestern areas, as shown in Fig. 6b.

Fig. 6
figure 6

a Wildfire risk zones based on MCDA-AHP; and b Pie chart showing wildfire risk zones percentage distribution

Full size image

3.3 Wildfire risk model performance evaluation

3.3.1 Wildfire risk model overlaid with historical fires

The final wildfire risk zones were overlaid with historical fire points and burned areas, as shown in Figs. 7a and 8a. The percentage distributions of historical fire and burned area were calculated per risk zone, as shown in Fig. 7b and Fig. 8b. Historical fires primarily occurred in the high-risk zone, with 40% occurring and 30% in the moderate-high-risk zone (Fig. 7b). The majority of the burned areas, 53%, were in the moderate-high-risk zone, followed by 15% in the low-risk zone (Fig. 8b).

Fig. 7
figure 7

a Wildfire risk zones overlaid with historical fire points; and b a Pie chart showing historical fire distribution (%) in the study area

Full size image
Fig. 8
figure 8

a Wildfire risk zones overlaid with burned area; and b a Pie chart showing burned area distribution (%) in the wildfire risk zones

Full size image

3.3.2 Wildfire risk model statistical tests performance

The model performance in predicting wildfire risk was assessed using the Confusion Matrix with historical fires as risk indicators per risk zone (Table 4). The overall and risk-specific statistics were subsequently analysed. Tables 5, 6, and 7 show the results of the Confusion Matrix, overall statistics, and risk statistics, respectively. Therefore, the model achieved an overall accuracy of 60% with a 50% Kappa’C (Table 6). The risk-specific statistic showed 100% accuracy in predicting moderate-high and high risks, followed by 88% accuracy for low risks (Table 7).

Table 5 Confusion matrix for wildfire risk model evaluation
Full size table
Table 6 Overall model statistics
Full size table
Table 7 Risk-specific statistics
Full size table

Furthermore, the relationship between historical fires and risk zones resulted in an MAE value of 4.28, an RMSE value of 4.84, and an R2 value of 0.85. Conversely, the relationship between burned area and risk zones was found to have an MAE value of 12.64, an RMSE value of 15.81, and an R2 value of 0.094, as shown in Fig. 9.

Fig. 9
figure 9

a Relationships between the wildfire risk zones; and b burned area (historical fires)

Full size image

3.4 Temporal pattern analysis

The analysis of historical fire occurrences and the temporal pattern of burned areas showed a weak positive correlation coefficient of 0.28 (Fig. 10a). The regression model p-value of 0.12 and R2 value of 0.08 indicate a weak relationship, indicating no statistical significance. The study found a negative relationship between historical fire occurrences and fire seasons, with a correlation coefficient of − 0.27, a p-value of 0.14 and an R2 value of 0.07 (Figs. 10b). However, the correlation coefficient for burned area and fire seasons was 0.30, indicating a weak positive relationship. The p-value was 0.10, and the R2 value was 0.09, with trends shown in Fig. 10b. The results show that historical fires were most frequent between 1992 and 1993, followed by 2000 and 2003, with the most burned areas occurring in 2017–2018 and 2018–2019, covering 64,099 and 96,063 ha, respectively.

Fig. 10
figure 10

a Time series showing the relationship between the burned area and fire seasons; and b historical fire points and fire seasons

Full size image

4 Discussion

The study assessed the spatial distribution of wildfire risk in Garden Route District (GRD), integrating human, topographic, and climatic factors using the MCDA-AHP to identify high-risk biodiversity hotspots. Historical fire data from 1991 to 2021 and Confusion Matrix were used to validate temporal wildfire trends and identify high and low-risk areas through burned area mapping using dNBR. This significantly contributes to wildfire risk assessment in a biodiversity hotspot.

4.1 Human factors influencing wildfire risk

The study considered human factors such as population density, LULC, distance from settlements, and roads. The results show that population density significantly influences fire risk, as an increase in density increases the likelihood of fire ignition. The study reveals a high population density in the wildland-urban interface (Fig. 3a), posing a risk of fire ignition, wildfire spread, damage, and fatalities due to flammable fynbos vegetation (GRDM-IDP 2022; Todorova et al. 2023). Wildfires are primarily caused by human activities like land use, population density, and distance from settlements, which directly increase the risk of wildfires (Maric et al. 2021; Todorova et al. 2023). The LULC map reveals fire points in vegetated and cultivated areas near built-up areas, indicating high-risk areas with LULC types (Fig. 3b), potentially igniting and spreading fires (Zhong et al. 2021; Afuye et al. 2024a). This is because fires can be utilised to revitalize farmland and grazing land in agricultural areas (de Nóbrega et al. 2018). The study revealed numerous point fires near settlements and roads, indicating a high risk of ignition and fire spread despite distance and road proximity (Figs. 3b and d). Shorter road distances, according to some scholars, improve accessibility, allowing firefighters to access fires promptly, and thereby reducing fire risk (Enoh et al. 2021). The final wildfire risk map revealed that past fires were primarily concentrated in high-risk zones characterized by agricultural land, vegetation, dense population, and denser settlements.

4.2 Topographic factors and wildfire propagation

Topographic factors significantly influence wildfire fuel and propagation, unlike human factors, but they do not necessarily affect ignition (Ajin et al. 2016, 2017; Güngöroğlu 2017; Lamat et al. 2021). The aspect map shows that fire points are concentrated on north and west-facing slopes, indicating high-risk areas (Fig. 4a). High solar radiation causes vegetation on slopes to dry up, acting as wildfire fuel, making it easier for wildfires to propagate (Kanga et al. 2017). The slope map revealed numerous fire points in mid-slope and upslope angles, indicating a high-risk zone, as fires move faster upslope than downslope (Fig. 4c). The slope significantly influences wildfire propagation, surpassing fuel generation due to its superior role in absorbing high solar radiation (Harrison et al. 2021). The results further show that areas with low elevation experience more fires than high-elevation areas (Fig. 4b). This connotes that dry conditions in low-elevation areas increase the risk of high risks due to the influence of elevation, temperature, and precipitation (Jensen et al. 2018). Dat & Son (2022) discovered that increased precipitation and decreased temperature in high-elevation areas reduce wildfire risk, possibly due to regional microclimate dynamics (Afuye et al. 2022, 2024b). Moreover, high elevations promote wildfire suppression rather than ignition due to low temperatures and less dry vegetation (Asori & Emmanuel 2020). The low fire point concentration can be attributed to the rare occurrence of lightning-ignited wildfires in high elevations (Calviño-Cancela et al. 2016).

4.3 Climatic factors and wildfire dynamics

Climate influences wildfire fuel, propagation, and intensity, which are also influenced by topography (Lamat et al. 2021). The dry summer climate and vegetative fuel load significantly contribute to wildfire ignition and spread (Mndela et al. 2023). The study area experienced low-to-moderate-risk class with low precipitation ranging from 19 to 31 mm, except for the northeastern and southeastern parts, which received high rainfall between 42 and 75 mm (Fig. 5a). However, the temperature fluctuated between 20 and 22 °C, leading to dry fuel and vegetation, potentially igniting wildfires and creating a high-risk zone in the northwestern and southern regions (Fig. 5b). Low precipitation and high temperatures in the study area could lead to dry conditions and increased vegetative fuel load, igniting and spreading wildfires (GRDM-IDP 2022). The study reveals that water vapour with high fire points ranging from 1.4 to 1.8 g/kg was found in moderate-to-high-risk zones, causing fewer fires (Fig. 5c). However, it also shows that the risk of wildfires in the study area is less affected by water vapour. The study revealed higher historical fire points in moderate-to-low-risk zones with mid-wind speeds ranging from 3.3 to 3.9 m/s and high-risk zones between 4.5 and 5.0 m/s, indicating more fire points in high-wind speed areas (Fig. 5d).

4.4 Spatial and temporal wildfire risk patterns

Based on the weight assignment and overlay analysis method, the wildfire risk map indicates that 26% of the southwestern parts are in high-risk zones. The northwestern region is the most severely affected, with 27% of factors affecting it. The findings indicate a higher likelihood of wildfires in this area spreading and intensifying. The model was validated with a low-risk zone of 19%, with 40% of historical fires in the high-risk zone, 9% in the low-risk zone, 53% in the moderate-high-risk zone, and 15% in the low-risk zone (Figs. 7a and b). The moderate-high-risk zone caused 30% of historical fire events, but only 12% of the study area was burned despite accounting for 53% of the most burned area. The study found that the high-risk zone only accounted for 11% of the burned area in the study area, indicating that the number of fire events is not proportional to the burned area (Figs. 8a and b). A study highlights a clear and significant correlation between historical wildfire events, burned areas, and fire seasons (Diaz 2022). The model demonstrated statistical performance with an overall accuracy of 60% and a Kappa’C of 50% (Table 6), suggesting a need for improvement (Yue et al. 2023). The model’s accuracy in predicting high and low-risk zones is commendable, with 88% accuracy for low-risk zones and 100% accuracy for moderate-high and high-risk zones. The model predictions of historical fire occurrences and wildfire risk zones showed low MAE and RMSE values with high R2 values. Therefore, the model predictions were found to have low error levels and were accurate in predicting fire occurrence in wildfire risk zones. This study confirms the findings of Johnson et al. (1998) and Diaz (2022) that there are no significant differences between burned areas and wildfire risk zones.

The temporal analysis revealed a weak positive correlation (r = 0.28) between historical fire occurrences and burned areas and a negative correlation (r = − 0.27) between historical fire occurrences and fire seasons (Figs. 10a and b). However, the burned area exhibited a weak positive correlation of 0.30 with fire seasons, despite a positive trend, which suggests that various factors contributed to wildfires and the burned areas. The study indicates that multiple factors contributed to wildfire occurrences and burned areas, with low R2 values between 0.07 and 0.09. The results show a weak correlation between observed trends and patterns, suggesting that recovery time after wildfire events significantly influences the connection between wildfire events and burned areas (Bright et al. 2019). Negative correlations between historical fire occurrences and fire seasons could impact the hydrological cycle, human and livestock, crops, and ecosystem functioning. The dry summer climate variation in Garden Route District (GRD) could potentially hinder socioeconomic and environmental sustainability, affecting land use planning. Wildfires in Mediterranean dry summer climates can significantly alter vegetation cover and plant growth, causing substantial economic losses (Afuye et al. 2021b; Aurelle et al. 2022). The GRD is highly susceptible to wildfires, causing significant economic losses due to its impact on forest cover, air quality, water security, and flooding risk. Wildfires in GRD cause water shortages, vegetation threats, and water quality issues, causing asset loss and loss of natural veld and grazing pastures, particularly during active burning years. Du Plessis and Van Zyl (2021), cited in Balocchi et al. (2022), discovered that post-wildfire peak and storm flow increases, while baseflow reduction is severe due to low moisture storage and hydrology, potentially threatened by increased wildfire frequency. The study emphasises the correlation between wildfire risk, burned areas, and historical fires in GRD biodiversity hotspots, highlighting the necessity for adaptive measures. Additionally, it identifies wildfire risk patterns, revealing ecological vulnerability and potential cascading effects on ecosystem functioning and resilience.

4.5 Policy implications on wildfire risk for forest management and land use planning

Wildfires are becoming more frequent and intense, posing significant challenges for forest management and land use planning. To effectively manage wildfire risk, a holistic approach combining policy, science, and community engagement is needed. This study identifies high-risk zones that can aid in prioritising wildfire prevention and management, enabling the implementation of fire suppression measures to control burns. Comprehensive policies involving forest health and restoration, strategic land use planning, public awareness campaigns, sustainable funding mechanisms, and advancing fire science are necessary. Addressing the wildfire paradox can help redefine fire management objectives to manage wildfires in a fire-adapted environment. Promoting networked governance, practical strategies, and community mapping exercises at the local level can enhance wildfire mitigation measures in high-risk biodiversity hotspot areas. The Garden Route District’s fire prevention policies should prioritise fostering fire-adapted communities, reducing wildfire user unit (WUI) losses, and promoting wildfire response and resilient landscapes. Governments can reduce wildfire risk by implementing actionable policies for safe fire practices, controlling burns, managing land, and requiring homeowners to clear 50 m for firebreaks and buffer zones.

4.6 Limitations and future prospects

Wildfire accuracy is influenced by changes in input datasets, uncertainties from data gaps, and temporal inconsistencies, which can affect risk assessments. The AHP model effectively captures various risk factors, but its subjective weighting may introduce biases into the risk prioritisation process. Future research should incorporate species-specific responses to landscape ecological dynamics using mechanistic models to predict future wildfire occurrence and behaviour. A larger dataset or field-based observations are needed to strengthen the connection between human, topographic, and climatic factors based on forest fire severity and risk zones. The study recommends further research on GIS-based wildland fire simulation to evaluate spatial patterns of wildfire behaviour and its effects on vegetation recovery, soil erosion, and air quality. The integration of advanced fire modelling tools like LANDFIRE, FARSITE, Wildland Fire Decision Support System, FireFamilyPlus, BEHAVE Plus, Fire Danger Rating Systems, and FlamMap will provide detailed simulations and risk analysis.

5 Conclusion

The wildfire risk distribution in Garden Route District was analysed using the GIS-based MCDA-AHP method and satellite-derived data, considering human, topographic, and climatic factors. The study highlights the significance of comprehensive wildfire risk assessment by identifying high-risk biodiversity hotspots validated through historical fire data and Confusion Matrix. Wildfire propagation in high-risk areas is influenced by factors like population density, wind speed, temperature, slope, LULC types, low elevation, and distance from settlements and roads. Topography and climate factors, such as water vapour, aspect, and high precipitation, have a lower impact on ignition risk. The study found that human activities like land use and distance from settlements significantly increase fire risk due to population density. Human activities largely impact environmental components, necessitating strategic land use policy, fire prevention, and public education to reduce wildfire risks. The study emphasises the need for adaptive management strategies in mitigating wildfire risk due to the influence of human activities. The findings of this study can help in assigning fire risk-based criterion weights to support decision-makers in wildfire prevention and management, especially in arid and semi-arid regions. The model demonstrates its reliability and suitability for assessing forest fire risk in biodiversity hotspots, including mapping burned areas. The study highlights the need for advanced models in estimating wildfire spread uncertainties, highlighting their intricate interplay in biodiversity hotspots for improved predictive accuracy.