Wildfire susceptibility mapping by incorporating damage proxy maps, differenced normalized burn Ratio, and deep learning algorithms based on sentinel-1/2 data: a case study on Maui Island, Hawaii

ABSTRACT

Climate change has contributed to the recent increase in wildfire occurrences, vegetation failures, human health risks, physical damage, and economic losses. Wildfire susceptibility mapping is an essential technique for assessing areas prone to wildfires. In this study, we proposed the combination of the damage proxy map (DPM) and differenced normalized burn ratio (dNBR) method to generate a precise wildfire inventory map and used it to predict areas susceptible to wildfire. The wildfire susceptibility maps were produced using frequency ratio (FR), convolutional neural network (CNN), and long short-term memory (LSTM)-based deep learning and their performances were compared. We implemented the proposed method on Maui Island, Hawaii, where wildfires frequently occur. We started the process by generating a wildfire inventory map from 2019 to 2023 based on the DPM method applied to Sentinel-1 synthetic aperture radar (SAR) data combined with a dNBR map retrieved from Sentinel-2 data. The wildfire inventory was randomly divided into a training dataset (70%) and a testing dataset (30%). Fifteen wildfire-related factors, including topographical, meteorological, land use, environmental, and anthropological factors, were selected to predict wildfires. The wildfire-related factors were selected by conducting study literature and considering spatial correlation analysis based on the FR method, information gain ratio analysis (IGR), and multicollinearity assessment using tolerance (TOL) and variance inflation factor (VIF) metrics. The level of susceptibility of an area to wildfire is divided into five, namely very high, high, moderate, low, and very low. The FR, CNN, and LSTM produced wildfire susceptibility maps with similar patterns, significantly influenced by land use and rainfall factors. The highly susceptible areas are located on gentle slopes covered by agricultural land and unhealthy vegetation, and these areas have low rainfall intensity but receive high levels of solar radiation. Meanwhile, areas with relatively low susceptibility occur in forests with high levels of wet canopy evaporation. The prediction results were evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), and the CNN performed slightly better than the FR and LSTM, with AUC values of 0.879, 0.877, and 0.870, respectively. Hence, the use of the CNN algorithm in the proposed method is appropriate, specifically for the study area. In addition, the susceptibility map provides information on susceptible areas on Maui Island, Hawaii, to increase public awareness.

KEYWORDS:

• Wildfire susceptibility map
• Maui Island
• damage proxy map (DPM)
• differenced normalized burn ratio (dNBR)
• deep learning

1. Introduction

The interaction among multiple environmental, climatic, and anthropogenic factors creates a dynamic wildfire risk landscape. Wildfire occurrences depend on the availability of resources or fuel load, ignition, and conditions favorable for ignition, which can originate from human or natural factors (Krawchuk et al. 2009). Human land management measures, such as agricultural land clearing and spatial landscape change, have caused biodiversity loss and ecosystem changes, locally increasing fuel and fire loads (Cai et al. 2021; Xu et al. 2022). Land use changes have globally contributed to climate change, causing increased temperatures and decreased precipitation levels that triggered prolonged droughts, leading to wildfire occurrences (Dupuy et al. 2020; Spracklen et al. 2009). In addition, changes in land use alter heat and water vapor surface fluxes, resulting in changes in thunderstorm patterns and altering the impact of wildfires (Pielke 2005). Moreover, human activities, such as land use changes and agricultural practices, play a significant role in altering fire regimes worldwide. Climate factors, including temperature, precipitation, and atmospheric circulation patterns, also influence the frequency and intensity of biomass-burning events (Chuvieco et al. 2021). Nevertheless, natural factors such as lightning ignition cannot be ignored as one of the leading causes of wildfires, and this factor will certainly increase the occurrence possibility under fire weather conditions (Krawchuk et al. 2009; Nur, Kim, and Lee 2022).

The natural conditions on the Island of Hawaii create dynamic weather patterns. The existence of an El Niño cyclone reduces rainfall, increases fuel originating from fire-prone vegetation, and leads to increased storm activity. Recently, Hawaii experienced dry conditions following an El Nino event, including moderate to severe drought and low precipitation accompanied by high temperatures (Frazier et al. 2019). These conditions consequently increased the number of wildfire occurrences over the last two decades, with more than 1,000 wildfires annually and a burned area of approximately 84.27 km². Dry vegetation, including grasslands and shrublands, was mostly burned, and the high building area density created the worst conditions for damage (Trauernicht et al. 2015). The destruction of buildings and productive lands might also cause the loss of jobs and homes, injuries, and even deaths (Bjånes, De La Fuente, and Mena 2021; Tonini et al. 2020). Approximately 24% of Hawaii is covered by dry and fire-prone grassland and shrubland areas. Hence, the combination of land cover, weather anomalies, and human factors facilitates rapid and widespread wildfires on Hawaii Island (Trauernicht et al. 2015). Therefore, establishing an inventory of wildfire occurrences and detecting areas susceptible to wildfire based on relevant factors are necessary to prevent damage, mitigate the associated risk, and design organized plans.

The inventory should contain wildfire record periods, locations, and amounts of historical wildfire occurrences. Recently, various sources have contributed to identifying, monitoring, and assessing wildfire inventories. Remote sensing data provided by different satellite sensors and platforms constitute one of the most valuable sources for identifying and tracking events. The Moderate Resolution Imaging Spectroradiometer (MODIS) sensor onboard the National Aeronautics and Space Administration (NASA)’s Terra and Aqua satellites (Schroeder et al. 2008; Ying et al. 2019), the Visible Infrared Imaging Radiometer Suite (VIIRS) onboard the Suomi National Polar-orbiting Partnership (Suomi NPP) and the National Oceanic and Atmospheric Administration − 20 (NOAA-20) satellites (Nur et al. 2023), and the Landsat and Sentinel-2 satellites have contributed to generating wildfire inventories (Hislop et al. 2018; Shaik, Giovanni, and Fusilli 2022). However, these optical images are limited by weather conditions, sunlight availability, and smoke occurrence, particularly during wildfires. Synthetic aperture radar (SAR) data can overcome these limitations. A near real-time progression of burn areas has been monitored by utilizing a time series of Sentinel-1 SAR data together with a deep learning model and successfully detecting wildfires with an overall accuracy of greater than 83% (Ban et al. 2020). The coherence difference between SAR data acquired on different dates provides a damage proxy map (DPM) representing wildfire occurrences. However, the study utilized official perimeters to select coherence differences representing wildfires manually (Nur, Kim, and Lee 2022).

The inventory map and related factors can be utilized to predict areas prone to wildfires. Various methods can be employed, including statistical, machine learning, and deep learning models. Statistical models include logistic regression (Preisler and Westerling 2007), frequency ratio (FR) (Heidari and Arfania 2022), and Bayesian models (Jaafari, Gholami, and Zenner 2017). However, statistical models exhibit limitations when managing complex relationships in data. Machine learning algorithms, such as the random forest (Bustillo Sánchez et al. 2021; Gholamnia et al. 2020), support vector machine, support vector regression (Piralilou et al. 2022), and XGBoost (Abujayyab et al. 2022) models, can overcome these limitations. Relevant studies represented that the capability of the machine learning model is better than the linear regression-based statistical method in predicting wildfire-prone areas (Iban and Sekertekin 2022; Shmuel and Heifetz 2022).

The machine learning models have shown promising capabilities in analyzing complex data in various cases. However, these models have a shallow architecture limited to extracting intricate data patterns and capturing information when complex and high-dimensional data is provided (C. Zhang et al. 2018). Deep learning methods addressed the problem by conducting prediction with feature extraction and producing significant improvement in the study of wildfire mapping (Muhammad, Ahmad, and Baik 2018). A convolutional neural network (CNN) (G. Zhang, Wang, and Lie 2021) and long short-term memory (LSTM) (Tin et al. 2019) are among the deep learning models that have been suggested to enhance the model accuracy in predicting areas susceptible to wildfires. The models have notably contributed to wildfire prediction by characterizing the nonlinear relationships among wildfire-related factors, such as temperature, drought, and historical events.

Predicting areas susceptible to wildfires requires a precise wildfire inventory, correct wildfire-related factors, and proper prediction models. For this reason, we proposed the combination of the DPM and dNBR techniques to produce precise wildfire inventories and investigated the utilization of deep learning models in mapping areas susceptible to wildfires based on the wildfire inventories obtained. The DPM method was applied to Sentinel-1 SAR data, and the dNBR map was generated using Sentinel-2 images acquired before and after nine wildfire events. Then, a deep learning method using the CNN and LSTM algorithms was applied to predict areas prone to wildfires using the dependent and independent variables. The inventory map was employed as a dependent variable, while fifteen wildfire-related factors were selected and served as independent variables. The model prediction performance was compared with a statistical technique-based frequency ratio that was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC). The best susceptibility map was obtained based on the comparison results. This study was implemented on Maui Island, Hawaii, which frequently experiences wildfires.

2. Study area

Maui Island is the second largest island of the group of islands called the Hawaiian Islands, located in the central Pacific Ocean. Maui Island is widely known for its lush landscape, which is mainly covered by forestland (37%), rangeland (32%), and agricultural areas (26%). Meteorologically, the island exhibits a tropical climate with dry and windy conditions during drought events. An official report indicated that moderate drought conditions occur in 20% of Maui, with severe drought conditions in 16% of the area (Hawaiʻi Climate Data Portal 2023). The natural climatic conditions of the Pacific Ocean cause complex weather patterns on Maui Island. El Niño and La Niña cycles cause periodic changes in Maui’s weather, in which a high rainfall intensity is associated with La Niña from the end to the beginning of the year, and drought, accompanied by hurricanes, is associated with El Niño from May to October. In addition, climate changes have been reported to affect rainfall reduction and temperature increase across Maui Island (Marris 2023).

The combination of vegetation-covered land, weather patterns, and climate change has led to wildfires in recent years. Figure 1 shows the wildfire perimeters reported by the Department of Natural Resources and Environmental Management (NREM) of the University of Hawaiʻi in Mānoa. At least 22 events occurred between 2019 and 2022, affecting areas of approximately 76.63 km² in 2019, approximately 4.3 km² in 2020, and 8.6 km² in 2022. These events mainly occurred between July and October during the dry season on Maui Island. Considering the characteristics of the island and the recorded wildfire occurrences, a wildfire susceptibility map should be obtained to better understand the primary factors, predict areas prone to wildfires, and prevent negative impacts in the future.

Figure 1. Land use in the study area and wildfire perimeters representing the burned area on Maui Island due to the recorded 2019–2023 events.

3. Materials and methods

This study mainly aimed to generate a wildfire inventory and susceptibility map. The main steps to gain the purposes included collecting satellite data, detecting wildfire location, collecting and selecting wildfire conditioning factors, predicting areas susceptible to wildfire, and validating the results. We utilized both Sentinel-1 and Sentinel-2 and assessed their accuracies individually and combined in detecting wildfire locations. The DPM method was employed on Sentinel-1 data, and the dNBR method was applied to Sentinel-2 data. The best result in detecting wildfires was considered as the wildfire inventory map. Selected wildfire-related factors were overlaid to wildfire inventory to extract valuable attributes and predict areas susceptible to wildfires based on frequency ratio, CNN, and LSTM model. Susceptibility levels were divided into five categories: very low, low, moderate, high, and very high. The performance of the models was evaluated based on the area under the ROC curve (AUC). The results were analyzed to find the influence of the factors on the occurrences of wildfires. Subchapters 3.1 to 3.4 and Figure 2 provide detailed explanations and workflow.

Figure 2. Diagram of the workflow for detecting wildfire locations and generating wildfire susceptibility maps.

3.1. Satellite data

SAR satellites produce images without cloud and smoke interference, which is beneficial during wildfire observations. Moreover, optical satellites provide images that can be obtained relatively easily, but they depend on weather conditions (Ban et al. 2020; Tariq et al. 2021). In this study, we used Sentinel-1 SAR data that employ a C-band with a 5.5-cm wavelength provided by the European Space Agency (ESA). Our study area is located in Path 124 and Frame 62 along the ascending flight direction and in Path 87 and Frame 521 along the descending direction. We selected nine wildfire events from 2019 to 2023 and collected 27 SAR images, with three scenes for each event, including two scenes before and one after the event (Table 1). We used single-look complex (SLC) data in the interferometric wide-swath (IW) mode under vertical-vertical (VV) polarization, in which the transmission and reception of radar signals occur along the vertical direction. The images were utilized to produce an interferometric coherence map of pre- and co-events for generating a damage proxy map (DPM) (Nur, Kim, and Lee 2022).

Table 1. Wildfire information and sentinel-1 data used to generate the DPM.

Wildfire Date	Pre-event		Co-event	Satellite Information
	1	2	3	Flight Direction	Path	Frame
8 August 2023	23 July 2023	4 August 2023	16 August 2023	Ascending	124	62
9 November 2022	20 October 2022	1 November 2022	13 November 2022	Ascending	124	62
31 July 2022	16 July 2022	28 July 2022	9 August 2022	Ascending	124	62
30 August 2020	10 August 2020	22 August 2020	3 September 2020	Descending	87	521
9 July 2020	29 June 2020	5 July 2020	17 July 2020	Descending	87	521
21 October 2019	30 September 2019	12 October 2020	24 October 2020	Ascending	124	62
2 October 2019	21 September 2019	27 September 2019	3 October 2019	Descending	87	521
1 August 2019	23 July 2019	29 July 2019	4 August 2019	Descending	87	521
12 July 2019	5 July 2019	11 July 2019	17 July 2019	Descending	87	521

In addition, we collected Sentinel-2 images of pre- and co-events provided by the ESA and calculated the median normalized burned ratio (NBR) map on the Google Earth Engine (GEE) platform (Table 2) (P. Liu et al. 2023; Sulova and Arsanjani 2021). In this study, we only used images of Maui Island with a cloud pixel percentage of less than 10%. NBR maps were generated to identify burned areas affected by a given event and then used to screen the DPMs.

Table 2. Wildfire information and sentinel-2 data to produce the median NBR map.

Wildfire Date	Pre-event	Co-event
8 August 2023	7 August 2023	15 August 2023
9 November 2022	22 October–8 November 2022	9 November–22 December 2022
31 July 2022	15 June–30 July 2022	1 August–8 September 2022
30 August 2020	1–29 August 2020	1–30 September 2020
9 July 2020	1 May–8 July 2020	10 July–10 August 2020
21 October 2019	25 August–20 October 2019	22 October–8 November 2019
2 October 2019	25 August–1 October 2019	22 October–8 November 2019
1 August 2019	15–31 July 2019	2 August–15 November 2019
12 July 2019	20 June–11 July 2019	15–31 July 2019

3.2. Wildfire inventory map

During wildfire inventory generation, we calculated interferometric coherence differences to obtain a damage proxy map (DPM). Interferometric coherence is a measure of the similarity of the phase difference between two SAR images acquired at different times over the same area. We initiated DPM generation by conducting image coregistration to match images with reference images with subpixel accuracy. The three coregistered images for each event were used to generate two interferometric coherence values: the coherence between a pair of images before the wildfire and the coherence between a pair of images spanning the wildfire (Han et al. 2021; Yun et al. 2015). The pre-event interferometric coherence (${\gamma }_{\mathit{pre}}$) was obtained by computing complex pixel values of two SLC images ($c_1$ and $c_2$) and their conjugate (${c_1}^{\ast }$ and ${c_2}^{\ast }$, respectively) based on Equation (1)(1) $\mathit{\gamma }=\frac{\left|⟨c_1{c_2}^{\ast }⟩\right|}{\sqrt{⟨}{c}_1{c_1}^{\ast }⟩⟨c_2{c_2}^{\ast }⟩},0\le \mathit{\gamma }\le 1$(1) (Tay et al. 2020). The same treatment was applied to a pair of SLCs spanning the event to calculate the co-event coherence (${\gamma }_{\mathit{co}}$). Then, the DPM was obtained by subtracting the ${\mathit{\gamma }}_{\mathrm{pre}}$ and ${\mathit{\gamma }}_{\mathrm{co}}$ maps, as expressed in Equation (2)(2) $\mathit{DPM}={\gamma }_{\mathit{pre}}-{\gamma }_{\mathit{co}}$(2) (Nur, Kim, and Lee 2022).

(1) $\mathit{\gamma }=\frac{\left|⟨c_1{c_2}^{\ast }⟩\right|}{\sqrt{⟨}{c}_1{c_1}^{\ast }⟩⟨c_2{c_2}^{\ast }⟩},0\le \mathit{\gamma }\le 1$(1)

(2) $\mathit{DPM}={\gamma }_{\mathit{pre}}-{\gamma }_{\mathit{co}}$(2)

The coherence value ranges from 0 to 1, with 0 indicating no coherence and 1 indicating perfect coherence (Nur, Kim, and Lee 2022). Generally, the vegetation area exhibits low coherence, while the built-up area exhibits high coherence. Damaged areas are defined as those areas exhibiting decorrelation or loss of coherence between two images and the resulting DPMs. The DPM value ranges from −1 to + 1, where negative values indicate changes uncorrelated with the wildfire, and positive values indicate wildfire-related changes. An appropriate threshold for positive DPM values was selected by considering burn scars in 3-m resolution PlanetScope images. Notably, the DPM shows not only coherence changes related to wildfires but also other changes at the Earth’s surface.

For this reason, we generated NBR maps to select the burned areas accurately. An NBR map was generated using the near-infrared (NIR) and shortwave infrared (SWIR) bands of Sentinel-2 images based on Equation (3)(3) $\mathit{NBR}=\frac{\mathit{NIR}-\mathit{SWIR}}{\mathit{NIR}+\mathit{SWIR}}$(3) .

(3) $\mathit{NBR}=\frac{\mathit{NIR}-\mathit{SWIR}}{\mathit{NIR}+\mathit{SWIR}}$(3)

To obtain an NBR map with less cloud coverage, we computed the median NBR on a pixel-by-pixel basis from several Sentinel-2 images before and after the event. The burned area was determined by subtracting the pre-event NBR value ($\mathit{NB}{R}_{\mathit{pre}}$) from the co-event NBR value ($\mathit{NB}{R}_{\mathit{co}}$), which is denoted as the differenced NBR (dNBR), as expressed in Equation (4)(4) $\mathit{dNBR}=\mathrm{NB}{\mathrm{R}}_{\mathit{co}}-\mathrm{NB}{\mathrm{R}}_{\mathit{pre}}$(4) .

(4) $\mathit{dNBR}=\mathrm{NB}{\mathrm{R}}_{\mathit{co}}-\mathrm{NB}{\mathrm{R}}_{\mathit{pre}}$(4)

The NBR value ranges from −1 to 1, with high values denoting healthy vegetation and low values denoting burned areas (Heidari and Arfania 2022). At the same time, the dNBR value ranges from −2 to + 2, with burned areas indicated by negative values. We manually adjusted the dNBR threshold to delineate the burned areas.

Moreover, we proposed to combine the DPM and dNBR methods by extracting DPMs below the selected dNBR threshold to cover the limitations of each method. The accuracy of wildfires detected by each method was compared based on confusion matrix analysis. Precision, recall, and F1-Score metrics are calculated using Equations 5–7 after well-detected or true positive (TP), underestimated or false negative (FN), and overestimated or false positive (FP) fire areas were obtained (Nemni et al. 2020; Shmuel and Heifetz 2023). Finally, the well-detected wildfires produced by the method with the best accuracy were used for wildfire susceptibility mapping. The workflow of our wildfire location detection method using DPMs combined with dNBR is shown in Figure 2.

(5) $\mathit{Precision}=\frac{\mathit{TP}}{\mathit{TP}+\mathit{FP}}$(5)

(6) $\mathit{Recall}=\frac{\mathit{TP}}{\mathit{TP}+\mathit{FN}}$(6)

(7) $\mathit{F}1-\mathit{score}=\frac{2\times \mathit{Precision}\times \mathit{Recall}}{\mathit{Precision}+\mathit{Recall}}$(7)

3.3. Modeling areas susceptible to wildfires

3.3.1. Wildfire-related factors

A total of 15 wildfire-related factors, categorized into topographical, meteorological, land use, environmental, and anthropological factors, as listed in Table 3 and Figure 3, were selected in this study to establish a model for mapping areas prone to wildfires. The topographical factors include the aspect (Figure 3a), slope (Figure 3b), topographic wetness index (TWI) (Figure 3c), and wind exposition index (Figure 3d). These factors were generated using a basic terrain analysis tool based on a digital elevation model (DEM) (Lee and Kim 2020). A DEM from the Shuttle Radar Topography Mission (SRTM) with a spatial resolution of 30 m provided by the United States Geological Survey (USGS) was used to obtain the topographical factors. The aspect is the direction of the slope surface that affects the variations in sunlight and moisture (Hakim et al. 2022). The slope degree is related to wildfires because this factor can influence fire distribution, where steep zones can exhibit faster fire distribution than gentle slopes (Nur et al. 2023). The TWI is a parameter that reflects moisture accumulation in a particular area and controls the hydrological process (Kadavi, Lee, and Lee 2018; Pourghasemi and Beheshtirad 2015). The wind exposition index indicates the area exposed to wind that can be estimated using the relative terrain aspect, wind direction, and horizon angle (Böhner and Antonić 2009). Negative values of the wind exposition index indicate the leeward direction, with positive values indicating the windward direction (Storey et al. 2020).

Figure 3. Wildfire-related factors: (a) aspect, (b) slope, (c) TWI, (d) wind exposition index, (e) solar radiation, (f) wet canopy evaporation, (g) windspeed, (h) rainfall, (i) land use, (j) NBR, (k) NDVI, (l) distance to streams, (m) distance to roads, (n) distance to buildings, and (o) distance to fire stations.

Figure 3. (Continued).

Table 3. Detailed information on the wildfire-related factors.

Category	Factor	Source
Topographical	Aspect	DEM SRTM, USGS
	Slope
	TWI
	Wind Exposition Index
Meteorological	Solar Radiation	Geography Department, University of Hawaiʻi (Giambelluca et al. 2023)
	Wet Canopy Evaporation
	Windspeed
	Rainfall	Hawaiʻi Climate Data Portal (Hawaiʻi Climate Data Portal 2023)
Land Use	Land Use	Office of Planning and Sustainable Development (Hawaiʻi State Office of Planning and Sustainable Development 2023)
Environmental	NBR	Sentinel-2, ESA
	NDVI	Sentinel-2, ESA
	Distance to Streams	Office of Planning and Sustainable Development (Hawaiʻi State Office of Planning and Sustainable Development 2023)
Anthropological	Distance to Roads
	Distance to Buildings
	Distance to Fire Stations

The meteorological factors include solar radiation (Figure 3e), wet canopy evaporation (Figure 3f), windspeed (Figure 3g), and rainfall (Figure 3h). Solar radiation influences soil moisture, temperature, and humidity, and it is reported as a factor triggering wildfires (Nur, Kim, and Lee 2022; Sayad, Mousannif, and Al Moatassime 2019; Tasie, Israel-Cookey, and Banyie 2018). Wet canopy evaporation refers to water evaporation from the wet surface of the Earth that is closely affected by the precipitation rate (Hadiwijaya et al. 2021). Windspeed is a wildfire-related factor because it is closely related to the resultant severity (Shakesby 2011). Moreover, rainfall is one of the important factors related to wildfire occurrence because it influences the fuel moisture content (Oliveira et al. 2014). Considering the variation of these meteorological factors over time, we obtained 12 monthly maps of solar radiation and wet canopy evaporation, respectively, from the Geography Department, University of Hawaiʻi, and averaged them into an average monthly map for this study. Meanwhile, the windspeed data is provided as annual hourly data. We obtained a mean of 24 hourly windspeed maps as a meteorological factor. The rainfall data is an average monthly rainfall from 2019 to July 2023 provided by the Hawaiʻi Climate Data Portal. The monthly map was produced using a climate-aided modified automatic kriging interpolation of a log-transformed sum of a monthly rainfall anomaly ratio $\frac{\mathit{observed}\mathit{rainfall}\left(\mathit{mm}\right)}{\mathit{mean}\mathit{monthly}\mathit{rainfall}\left(\mathit{mm}\right)}$. This kriging process used 70 unique station locations within Maui County (Maui, Lanai, Molokai, and Kahoolawe) and a total of a month of recorded or estimated rainfall (mm). Monthly rainfall maps from 2019 to July 2023 were averaged and used as the rainfall factor in this study.

A land use map (Figure 3i) is also considered a wildfire-related factor. Various land uses exhibit different characteristics. The severity and spread of wildfires are related to the land coverage. Differences in land use landscape features affect soil moisture and the provision of fire ignition (Nur, Kim, and Lee 2022). Therefore, land use maps constitute one factor that is most considered in wildfire occurrences.

The environmental factors include the NBR (Figure 3j), normalized difference vegetation index (NDVI) (Figure 3k), and distance to streams (Figure 3l). An NBR map separates healthy vegetation and burned areas by exploiting the spectral responses of the NIR and SWIR bands (Heidari and Arfania 2022). At the same time, in the NDVI map, different colors are assigned to describe vegetation presence, condition, and health (Nhongo et al. 2019). In this study, the median NBR and NDVI values from 2019 to 2023 were calculated to prevent bias resulting from the reduction in vegetation following a wildfire.

The anthropological factors include the distance to roads (Figure 3m), distance to buildings (Figure 3n), and distance to fire stations (Figure 3o). Regarding the distance to roads, Narayanaraj and Wimberly (2012) found that some wildfires were closely correlated with fire ignition caused by humans close to roads (Narayanaraj and Wimberly 2012). The distance of a given area to buildings should be analyzed because the increase in the number of buildings can increase the number of wildfire occurrences. The wildfire severity and spread are correlated with building characteristics, such as the structure and material type, inclination of the ground, and surrounding environment (Papathoma-Köhle et al. 2022). The location of fire stations is considered a wildfire-related factor because it is important in minimizing widespread damage caused by wildfires (Savsar 2014). In modeling areas prone to wildfires, all wildfire-related factors were adopted as independent variables, while the wildfire occurrence location was employed as a dependent variable.

3.3.2. Feature selection of the wildfire-related factors

Various factors can influence wildfires, and the correlation between these factors and wildfire occurrence must be assessed in wildfire probability modeling. However, the utilization of all factors can influence the model prediction performance, causing a decrease in accuracy. Relatively essential factors must be investigated to select relevant factors that lead to wildfires to ensure the model’s performance. In this research, we employed information gain ratio (IGR) analysis to investigate the factors, with the IGR value ranging from 0 to 1. Higher IGR values indicate greater model prediction capabilities, while lower values indicate the opposite. Furthermore, multicollinearity was assessed using the tolerance (TOL) and variance inflation factor (VIF) metrics to identify essential factors associated with wildfire occurrences. TOL and VIF can be determined based on the correlation between a certain wildfire-related factor and other factors ($R^2$), which can be computed using Equations (8) and (9), respectively.

(8) $\mathit{TOL}=1-R^2$(8)

(9) $\mathit{VIF}=\frac{1}{1-R^2}$(9)

The frequency ratio (FR) analysis method is a statistical method to analyze the spatial correlation between the location of wildfires and classes of related factors (Lee and Rezaie 2021). The FR method was used to infer the probability of wildfire occurrence by assuming that conditions will remain unchanged (Lee and Talib 2005). FR values were calculated based on Equation (10)(10) $\mathit{F}{R}_{\mathit{ij}}=\frac{{\mathit{W}}_{\mathit{ij}}/N_w}{F_{\mathit{ij}}/N_F}=\frac{\mathrm{\%}\mathit{Pixels}\mathit{of}\mathit{wildfires}}{\mathrm{\%}\mathit{Pixels}\mathit{of}\mathit{class}}$(10) by determining the ratio between the percentage of wildfire location pixels and the percentage of factor class pixels (Lee and Rezaie 2021; I. Park and Lee 2014).

(10) $\mathit{F}{R}_{\mathit{ij}}=\frac{{\mathit{W}}_{\mathit{ij}}/N_w}{F_{\mathit{ij}}/N_F}=\frac{\mathrm{\%}\mathit{Pixels}\mathit{of}\mathit{wildfires}}{\mathrm{\%}\mathit{Pixels}\mathit{of}\mathit{class}}$(10)

where $\mathit{F}{R}_{\mathit{ij}}$ denotes the FR value of the $\mathit{i}$-th class of the $\mathit{j}$-th factor, $W_{\mathit{ij}}$ is the number of wildfire pixels in the $\mathit{i}$-th class of the $\mathit{j}$-th factor, $N_W$ is the number of wildfire pixels, $F_{\mathit{ij}}$ denotes pixel count in the $\mathit{i}$-th class and the $\mathit{j}$-th factor, and $N_F$ represents the cumulative pixel count for each factor.

3.3.3. Convolutional neural network (CNN)

The CNN is generally designed for processing and analyzing visual data and involves the application of convolution operations to extract detailed features from the input data automatically. The CNN layers include the input layer, convolutional layer, pooling layer, rectified linear unit (ReLU) function activation layer, fully connected layer, and output layer, where each layer fulfills a specific role in feature extraction, pattern recognition, and classification (Panahi et al. 2020). Images are input into the input layer, the convolutional layer extracts feature from the input layer and learns the connection between the features based on Equation 11 (Khosravi et al. 2020), and the activation function, namely, ReLU, is applied to introduce nonlinearity into the model (Panahi et al. 2020).

(11) $y_j=\mathit{f}\left({\sum }_i{k}_{\mathit{ij}}\ast x_i+b_j\right).$(11)

where $y_j$ is the output of the $\mathit{j}$th convolution layer, $\mathit{f}$ denotes the activation function (ReLU), $k_{\mathit{ij}}$ denotes the convolution kernel for the $\mathit{i}$th input feature map $x_i$, and $b_j$ denotes the bias (Zhang et al. 2022). The pooling layer reduces the number of parameters to improve the model performance without losing essential information using the maximum pooling or average pooling technique (X. Zhang, Wu, and Li 2021). The fully connected layer serves to connect all neurons of the previous layer and enables classification to gain the final decision (Lee and Rezaie 2021). The output layer provides the probability distribution for the various classes in the classification task.

3.3.4. Long short-term memory (LSTM)

The LSTM is an advanced recurrent neural network (RNN) that can be used to process and model sequential datasets. The LSTM addresses the long-term dependency limitation of RNNs by providing a forget gate to remove or forget unnecessary information from the previous cell. The LSTM structure includes a memory cell, and each cell contains three gates comprising the input gate ($i_t$), forget gate ($f_t$), and output gate ($o_t$). These gates are responsible for removing or adding information to the cell (Tin et al. 2019; Yang, Pan, and Tao 2017). The input gate controls the information in the current input that will be stored in the memory cell; the forget gate determines the information that should be kept or removed; and the output gate organizes the information received from the memory cell that must be used as output at the current step to be passed on to the next layer. At time $\mathit{t}$, an LSTM cell needs the input $x_t$ and output $y_{\mathit{t}-1}$ of the previous cell. If the cell input state is denoted as ${\tilde{C}}_{\mathit{t}}$, the cell output state is $C_t$, and the previous cell state is $C_{\mathit{t}-1}$. $C_t$ and $y_t$, which can be calculated by Equations 12 and 13, respectively, are transmitted to the next cell.

(12) $C_t=f_t\ast C_{\mathit{t}-1}+i_t\ast {\tilde{C}}_{\mathit{t}}$(12)

(13) $y_t=o_t\ast tanh(C_t)$(13)

where $i_t$, $f_t$, $o_t$, and ${\tilde{C}}_{\mathit{t}}$ can be calculated using Equations 14–17 following the calculation rules for the sigmoid function ($\mathit{\sigma }$) and the hyperbolic tangent function $\tanh$, as well as considering weighted matrices ($\mathit{w}$) and bias vectors ($\mathit{b}$) (Tin et al. 2019).

(14) $i_t=\mathit{\sigma }\left(w_i{x}_t+w_i{y}_{\mathit{t}-1}+b_i\right)$(14)

(15) $f_t=\mathit{\sigma }\left(w_f{x}_t+w_f{y}_{\mathit{t}-1}+b_f\right)$(15)

(16) $o_t=\mathit{\sigma }\left(w_o{x}_t+w_o{y}_{\mathit{t}-1}+b_o\right)$(16)

(17) ${\tilde{C}}_{\mathit{t}}=tanh\left(w_C{x}_t+w_C{y}_{\mathit{t}-1}+b_C\right)$(17)

3.4. Evaluation of the Model performance

In this study, the performance of the models in predicting the areas susceptible to wildfires was evaluated using receiver operating characteristic (ROC) curve analysis. In the ROC curve, the value of 1-the specificity or the false-positive rate is plotted on the x-axis, and the sensitivity or true positive rate is plotted on the y-axis (Al-Abadi et al. 2017). Then, the area under the ROC curve (AUC) can be calculated as a model evaluation metric. The AUC has been commonly used in evaluating the prediction accuracy of machine learning models for various cases, including wildfire modeling (Bustillo Sánchez et al. 2021; Nur et al. 2023). AUC values were obtained on the testing dataset not used in the model training process. The AUC values range from 0.5 to 1. The highest accuracy is denoted by a value of 1, while poor accuracy is denoted by a value of 0.5. In machine learning implementation studies, acceptable AUC values are generally higher than 0.8 (Nur et al. 2023; S. J. Park et al. 2018).

4. Results

4.1. Wildfire inventory map

Maui Island, Hawaii, experiences wildfires during the dry season, influenced by an increase in the drought level, rise in the temperature, and change in the weather pattern due to climate changes. Wildfires have occurred in areas covered mainly by flammable vegetation. Changes in the pre- and co-wildfire interferometric coherence values can be measured by SAR images without weather disturbances. At the same time, a dNBR map obtained using NIR and SWIR bands of optical images can also detect burned areas during wildfires. A combination of DPM and dNBR methods was investigated to provide a better wildfire inventory map. Table 4 shows that combining DPM and dNBR is the best method for producing wildfire inventories. The DPMs were geocoded according to the SRTM DEM with a spatial resolution of 30 m. We chose a coherence loss threshold for DPM values ranging from 0.1 to 0.6 and varying from more than −0.2 for dNBR values which represent areas damaged by wildfires, as indicated by yellow to red pixels.

Table 4. Comparison of accuracy of each method used in generating wildfire inventories.

Method	TP	FP	FN	TN	Precision	Recall	F1-score
DPM	2239	1010	861	46307	0.689	0.722	0.705
dNBR	1523	1726	145	47023	0.469	0.913	0.619
DPM+dNBR	2624	625	979	46189	0.808	0.728	0.766

The DPM method can produce better well-detected (TP) wildfires with a lower overestimated section (FP) compared to dNBR. At the same time, the dNBR method is more successful in reducing the underestimated section by producing a lower FN value than DPM. However, both methods failed to produce more than 70% precision values. By combining the DPM and dNBR methods and selecting only the wildfire detected by both methods as inventories, the limitations of each method can be solved. The comparison concluded that combining DPM and dNBR is the best method for creating an inventory map of wildfire, especially on the island of Maui. Figure 4 shows the areas damaged by severe wildfires from 2019 to 2023 based on DPM combined with dNBR. We overlaid the wildfire inventory results onto the natural color Sentinel-2 image, where the transparent pixels indicated no damage due to wildfires. The detected wildfires remained within the official reported wildfire perimeters. Regarding the 2019 to 2023 events, wildfires occurred from the central to northwestern parts of Maui Island.

Figure 4. Wildfire inventory map generated by combining the DPM and dNBR methods using sentinel-1 and sentinel-2 data.

A total of 9 wildfires were observed, and we determined a total of 73,818 wildfire pixels detected by both DPM and dNBR methods with a cell size of 30 m. These pixels were used to generate a wildfire susceptibility map of Maui Island using deep learning approaches by dividing them into a training dataset (70%; 51,673) and a testing dataset (30%; 22,145). Considering the deep learning method requirements, we sampled the same number of nonwildfire data and divided them into training (70%) and testing datasets (30%), similar to the wildfire data (Fadhillah et al. 2022; Tien et al. 2017). The nonwildfire data were randomly sampled from wildfire-free areas with a low wildfire occurrence probability, as determined by frequency ratio analysis (Hakim et al. 2022). The training data was used to train the models, while the testing data were employed to evaluate the model performance.

4.2. Assessment of the wildfire-important factors

The selection of wildfire-related factors is essential to ensure the model’s performance in generating susceptibility maps. Therefore, factors unrelated to wildfire occurrences should be eliminated to provide more accurate results. In this study, the wildfire-related factors were selected based on information gain ratio (IGR) analysis. The IGR value ranges from 0 to 1, where a value of 0 indicates that the factor is unrelated to wildfire occurrence, while a value close to 1 indicates that the factor is highly related to wildfire occurrence (Feng et al. 2022). Based on IGR analysis, 15 critical factors (with IGR > 0) for wildfire occurrence were selected, with the value of each factor provided in Table 5. Land use was the most highly correlated factor for wildfire occurrence, yielding the highest IGR value of 0.47, followed by rainfall (an IGR value of 0.46), distance to roads (an IGR value of 0.38), and NBR (an IGR value of 0.36). At the same time, the distance to streams exhibited the lowest IGR value (0.02).

Table 5. The relative importance of the wildfire-related factors and multicollinearity analysis results using the IGR, TOL, and VIF metrics.

Category	Factor	IGR	TOL	VIF
Topography	Aspect	0.06	0.88	1.13
	Slope	0.25	1.39	0.28
	TWI	0.03	1.14	0.13
Meteorological	Solar Radiation	0.19	1.40	0.28
	Wet Canopy Evaporation	0.27	2.24	0.55
	Wind Exposition Index	0.21	1.85	0.46
	Windspeed	0.02	1.44	0.31
	Rainfall	0.46	3.15	0.68
Land Use	Land Use	0.47	1.48	0.32
Environmental	NBR	0.36	3.43	0.71
	NDVI	0.27	2.71	0.63
	Distance to Streams	0.02	1.05	0.04
Anthropological	Distance to Roads	0.38	1.83	0.45
	Distance to Buildings	0.17	1.33	0.25
	Distance to Fire Stations	0.25	1.20	0.16

In addition, the multicollinearity of the wildfire-related factors should be assessed to ensure no highly correlated factors. In this study, we employed the tolerance (TOL) and variance inflation factor (VIF) metrics to evaluate the factor multicollinearity problem. A factor with VIF > 5 or TOL < 0.2 indicates a multicollinearity problem and should be removed before further analysis (Chen et al. 2018). The multicollinearity of all wildfire-related factors in this study met the acceptable threshold, with TOL and VIF values ranging from 0.88 to 3.43 and 0.04 to 1.13, respectively, as listed in Table 5. The result showed that the NBR exhibited the highest TOL value (3.43), followed by rainfall (3.15), NDVI (2.71), and wet canopy evaporation (2.24). Moreover, the lowest TOL value was obtained for the aspect (0.88). In VIF analysis, the maximum value of 1.13 was observed for the aspect, which is still lower than the acceptable threshold.

The spatial relationship between wildfire occurrence and related factors was analyzed using the FR method. The FR method provides the weight for each factor by examining the wildfire occurrence frequency in each class (Acharya, Yang, and Lee 2017). The FR between the wildfire occurrences on Maui Island, Hawaii, and related factors was calculated, as summarized in Table 6. A class of related factors with an FR value greater than 1 indicates a high correlation with wildfire occurrence. Table 6 shows that the recorded wildfires were concentrated on terrain surfaces with south-, southwest-, and west-facing aspects, slopes lower than 8°, and areas with TWI values ranging from 5.43–6.64 and 9.8–22.12. Moreover, areas classified as receiving solar radiation higher than 202.74 W/m² were more notably correlated with wildfire occurrence. The FR value of wet canopy evaporation showed that areas with low monthly water evaporation, including the 2.21–147.99 and 148–232.75 mm classes, were more closely correlated with wildfire occurrence. Wildfires also frequently occurred in areas exposed to low winds, as indicated by wind exposition index values between 0.76 and 1.04 and high monthly wind speeds ranging from 2.32 to 6.83 m/s. The rainfall factor showed that the two lowest rainfall classes were highly correlated with wildfire occurrence. The FR values of the land use factor indicated that areas covered by agricultural and and rangeland were highly correlated with wildfire occurrence. Moreover, the NBR and NDVI classes suggested that wildfires frequently occurred in low-vegetation areas. In addition, areas within 1,610 m from roads, 188 to 935 m from streams, and 4,262 to 6,799 m from fire stations, and all areas except those more than 4,030 m from buildings were more closely correlated with historical wildfires on Maui Island.

Table 6. Frequency ratio scores representing the relationship between wildfire occurrence and related factors.

Factor	Class	Percentage of each class	Percentage of wildfires	FR
Aspect	Flat	5.93	1.76	0.30
	North	11.50	5.98	0.52
	Northeast	8.94	5.94	0.66
	East	10.62	6.98	0.66
	Southeast	13.35	10.64	0.80
	South	9.54	17.90	1.88
	Southwest	14.75	24.06	1.63
	West	16.83	18.40	1.09
	Northwest	8.55	8.35	0.98
Slope (°)	0–4.03	18.29	57.00	3.12
	4.04–8.07	21.53	24.52	1.14
	80.71–12.41	20.26	6.28	0.31
	12.42–19.86	20.30	6.34	0.31
	19.87–79.12	19.63	5.86	0.30
TWI	1.45–5.42	18.74	10.14	0.54
	5.43–6.64	20.52	24.43	1.19
	6.65–8.02	20.66	20.06	0.97
	8.03–9.79	20.46	17.35	0.85
	9.8–22.12	19.62	28.02	1.43
Solar radiation (W/m^2)	129.96–187.67	19.92	0.26	0.01
	187.68–202.73	19.85	4.87	0.25
	202.74–217.79	20.09	22.44	1.12
	217.8–234.1	19.82	50.77	2.56
	234.11–289.94	20.31	21.67	1.07
Wet canopy evaporation (mm)	2.21–147.99	19.13	35.03	1.83
	148–232.75	20.22	42.87	2.12
	232.76–314.11	20.72	15.77	0.76
	314.12–459.89	20.01	5.86	0.29
	459.9–866.71	19.92	0.46	0.02
Wind exposition index	0.76–0.99	19.77	48.22	2.44
	1.00–1.04	20.28	32.26	1.59
	1.05–1.06	20.17	9.17	0.45
	1.07–1.10	19.92	5.45	0.27
	1.11–1.34	19.86	4.91	0.25
Windspeed (m/s)	0.43–1.84	19.49	7.70	0.40
	1.85–2.31	20.75	19.09	0.92
	2.32–2.86	19.63	21.53	1.10
	2.87–3.42	20.60	27.42	1.33
	3.43–6.83	19.53	24.26	1.24
Rainfall (mm)	17.89–33.82	19.45	63.88	3.28
	33.83–57.57	21.56	28.59	1.33
	57.58–107.71	20.57	7.31	0.36
	107.72–226.47	19.54	0.21	0.01
	226.48–690.94	18.88	0	0
Land use	Built-up land	2.94	2.60	0.88
	Agricultural land	25.74	59.69	2.32
	Rangeland	31.46	33.65	1.07
	Forestland	36.10	1.75	0.05
	Water body	0.15	0.24	1.58
	Wetland	0.13	0	0.00
	Barren land	3.47	2.06	0.59
NBR	−0.96–0.06	19.05	51.99	2.73
	0.07–0.25	20.72	38.21	1.84
	0.26–0.45	20.22	8.48	0.42
	0.46–0.62	19.35	1.24	0.06
	0.63–0.89	20.66	0.08	0
NDVI	−0.89–0.29	0.1951	0.3261	1.67
	0.3–0.44	0.2031	0.4778	2.35
	0.45–0.66	0.2063	0.1691	0.82
	0.67–0.82	0.1973	0.0252	0.13
	0.83–1	0.1982	0.0019	0.01
Distance to streams (m)	0–62.4	15.26	12.50	0.82
	62.5–187	22.60	19.70	0.87
	188–374	21.65	26.29	1.21
	375–935	21.51	34.60	1.61
	936–7,950	18.99	6.92	0.36
Distance to roads (m)	0–179	19.86	25.11	1.26
	180–681	21.68	39.79	1.84
	682–1,610	20.14	27.99	1.39
	1,620–3,260	19.43	7.09	0.37
	3,270–9,130	18.89	0.01	0
Distance to buildings (m)	0–792	19.55	20.13	1.03
	793–1,670	20.12	23.55	1.17
	1,680–2,730	20.55	24.76	1.21
	2,740–4,020	19.80	20.72	1.05
	4,030–7,480	19.98	10.83	0.54
Distance to fire stations (m)	0–4,261	19.74	19.95	1.01
	4,262–6,799	20.28	41.00	2.02
	6,800–9,610	19.99	30.67	1.53
	9,611–13,600	20.03	8.37	0.42
	13,610–23,120	19.96	0.01	0

4.3. Wildfire susceptibility map

Wildfire susceptibility maps were generated using the training dataset of 70% of the total wildfire inventory data. The FR assessed the likelihood of wildfires occurring in specific areas statistically, and areas susceptible to wildfires were mapped. The FR values of all factors were obtained by calculating the cross-tabulated area between the training data and the 15 wildfire-related factors. Thus, two deep learning algorithms, namely, the CNN and LSTM, were also employed. Areas susceptible to wildfires were classified into five categories: very low, low, moderate, high, and very high. This classification was determined based on the quantile technique that aims to divide the dataset into equal-sized classes based on the data distribution. Figures 5(a–c) show the wildfire susceptibility maps produced by the FR, CNN, and LSTM, respectively. The results showed that areas with higher susceptibility occur at the center of Maui and along the northwest coast of the island where Lahaina city is located. The color pattern from green to red indicates the very low to very high susceptibility classes. In general, the pattern of the wildfire susceptibility maps generally followed that of the land use and rainfall maps, Figure 3i and h. Notable differences were observed in built-up land (marked by black boxes), where the FR categorized the area as having a very high level of vulnerability, CNN categorized the area as exhibiting a high to very high susceptibility level, while the LSTM categorized it as exhibiting a moderate to high susceptibility level.

Figure 5. Wildfire susceptibility maps of Maui Island, Hawaii, based on the (a) FR, (b) CNN, and (c) LSTM algorithms.

By comparing the three wildfire susceptibility maps, the percentage distribution of pixels for each susceptibility category was analyzed, as shown in Figure 6. The FR-based susceptibility map categorized 26.38% of pixels of the study area at very low, 11.44% low, 11.43% moderate, 13.94% high, and 36.81% very high susceptibility levels. The susceptibility map produced by the CNN model predicted that approximately 19.21%, 20.46%, 14.82%, 12.79%, and 32.72% of the study area exhibited very low, low, moderate, high, and very high susceptibility levels, respectively. At the same time, the LSTM model indicated that approximately 20.73%, 21.23%, 16.46%, 13.42%, and 28.16% of the study area exhibited very low, low, moderate, high, and very high susceptibility levels, respectively. The percentage of pixels categorized as very low, high, and very high levels by FR is the highest compared to the two deep learning models. Meanwhile, CNN estimated a lower percentage of the pixel distribution than the LSTM, except for areas categorized as exhibiting a very high susceptibility level. The performance of the FR, CNN, and LSTM in predicting susceptible areas was validated using the AUC of the ROC curve on the testing dataset. The wildfire susceptibility map generated by the FR produced a susceptibility map with an AUC value of 0.877. At the same time, the CNN showed an AUC value of 0.879, and the LSTM-generated map indicated an AUC of 0.870 (Figure 7).

Figure 6. Comparison of the pixel percentages between the wildfire susceptibility classes produced by the FR, CNN, and LSTM.

Figure 7. Validation of the FR, CNN, and LSTM models based on AUC of ROC curve analysis.

5. Discussion

The availability of a wildfire inventory is essential for predicting areas prone to wildfires, which produces the susceptibility map. In this study, we used SAR data and generated DPMs to obtain a wildfire inventory. However, noise emerged in the result because the DPM not only captured coherence changes related to wildfires but also captured those related to other sources, such as construction, harvesting, and building collapse due to earthquakes, typhoons, and other phenomena (Han et al. 2021). We proposed a combining technique for removing noise unrelated to wildfire occurrence and increasing the wildfire detected using the dNBR map, which significantly reduced the noise level and yielded results that followed the official wildfire perimeter data provided by the Hawaii government. The proposed method showed areas damaged by wildfires and could be used to represent wildfire occurrence. At the same time, nondamaged areas were defined, and a wildfire inventory map of Maui Island was obtained. However, the proposed method has limitations when detecting wildfires in a small area due to the resolution of the data used in this study. A higher resolution image is needed to improve wildfire detection ability and produce a more detailed inventory map. Besides, satellites with thermal sensors may improve the ability to detect rapidly when a wildfire occurs.

The wildfire susceptibility maps of Maui Island produced by the FR, CNN, and LSTM exhibited similar patterns of susceptible areas. Based on IGR analysis, land use and rainfall are the most important factors of wildfire occurrence because they are related to the provision of fuel and the moisture and drought levels (Verde and Zêzere 2010). In addition, FR analysis showed that the wildfire susceptibility maps generally showed a high to very high susceptibility level in areas with flat to gentle slopes covered by agricultural land (Table 6). In addition, the highly susceptible areas exhibit a low annual rainfall intensity but receive high solar radiation. The slope and solar radiation are closely correlated with wildfire occurrence in addition to land use and rainfall. In the study area, gentle slopes were covered by unhealthy vegetation, as represented by the NBR and NDVI (G. Zhang, Wang, and Lie 2021). The accumulation of unhealthy vegetation could influence fuel availability, ignition threshold, and fire spread, which increases wildfire susceptibility (Piralilou et al. 2022). Moreover, solar radiation reflects sunlight and its intensity and notably influences the temperature. Conversely, solar radiation has an inverse relationship with relative humidity, where an increase in solar radiation causes a decrease in relative humidity, representing a decrease in rainfall (Tasie, Israel-Cookey, and Banyie 2018). Areas with fuels exposed to intense solar radiation are more likely to experience ignition due to the increased temperature and decreased humidity (Y. Liu, Goodrick, and Heilman 2014). Conforming with this consideration, land areas covered by forests with high wet canopy evaporation levels that are rarely exposed to sunlight indicated a low susceptibility (Table 6).

In this study, we relied on the ability of deep learning models, i.e. the CNN and LSTM algorithms, to predict areas prone to wildfires for producing susceptibility maps. The ability of deep learning models to yield predictions depends on the accuracy of the input data, and inaccurate predictions can result from incorrect inventory data (Lee and Rezaie 2021). The validation showed that the AUC value of the CNN was 0.879, the FR was 0.877, and the LSTM was 0.870. The AUC values are higher than 0.7, denoting a suitable performance of the models in predicting potential wildfire areas (Fadhillah et al. 2022). The percentage of historical wildfire locations from 2019 to 2023 was mainly classified into the very high class (Figure 6); therefore, the model performance is acceptable. Moreover, the CNN performed slightly better than the FR and LSTM in predicting susceptible areas, agreeing with the modeling study of susceptible areas conducted by Kavzoglu et al. (2021). The CNN excels in processing spatial information and extracting relevant features from integrated and diverse data sources, such as creating wildfire susceptibility maps (G. Zhang, Wang, and Liu 2021, 2019). The model performance must be assessed to provide accurate wildfire susceptibility maps. Even though the three models produced similar results, notable differences can be observed in the black box in Figure 5, where the FR model overestimated the susceptibility level, while the LSTM model underestimated the level relative to the CNN. Moreover, the black box areas mostly comprise built-up land (Figure 3i), closely correlated with human activities. Underestimating the susceptibility level could lead to inappropriately planned prevention and mitigation actions. Furthermore, the models are effectively applied in Maui Island, which is unique in climate and location. Nevertheless, the same characteristics of another study area should be investigated to validate the performance of our models as a further work.

6. Conclusion

Wildfire inventory maps of Maui Island, Hawaii, were generated by applying the DPM technique to SAR data, dNBR to optical images, and their combination. The dNBR maps produced from Sentinel-2 images were combined with DPM to remove noise unrelated to wildfires in the DPMs and get a more precise inventory map. A wildfire inventory was compiled, and related factors were chosen to generate wildfire susceptibility maps using the FR, CNN, and LSTM algorithms. Fifteen wildfire-related factors were selected based on spatial correlation and multicollinearity assessment. Each model was validated by calculating the AUC value of the ROC curve. Model validation in predicting areas susceptible to wildfires showed that the AUC value of the CNN (0.879) was slightly higher than that of the FR (0.877) and the LSTM (0.870). The wildfire susceptibility maps produced by the FR, CNN, and LSTM were categorized into very low, low, moderate, high, and very high susceptibility levels, and the results indicated similar patterns. IGR analysis showed that land use and rainfall are the most important factors influencing wildfire occurrences on Maui Island. In addition, FR analysis revealed relatively highly susceptible areas with gentle slopes covered with agricultural land and unhealthy vegetation, and these areas exhibit a low rainfall intensity but receive high solar radiation. Moreover, the relatively low-susceptibility areas occurred in forests with high wet canopy evaporation levels. Nevertheless, notable differences were observed. The LSTM underestimated the susceptibility in built-up land, indicating a moderate to high susceptibility level. Consequently, inappropriate prevention and mitigation actions could potentially occur. Therefore, the model performance should be a concern when areas prone to wildfires are being predicted. Furthermore, wildfire detection and susceptibility map generation should be enhanced in further studies by evaluating satellite data obtained with higher resolution and having a thermal sensor and validating in another study area. The results of this research could provide information on areas susceptible to wildfires, increase public awareness, and help concerned parties design appropriate prevention and mitigation actions.

Acknowledgments

The authors are grateful for the constructive comments and suggestions provided by the anonymous reviewers and editors, which have improved the quality of the manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The Sentinel-1 data that support the findings of this study are openly available in the NASA Earth Observation Data at https://search.asf.alaska.edu/. The DEM SRTM data is provided by the United States Geological Survey (USGS) at https://earthexplorer.usgs.gov/. The Sentinel-2 data was obtained from Google Earth Engine. The wildfire perimeters were digitized from PlanetScope images that are freely accessed at https://www.planet.com/for Education and Research Program members. The solar radiation, wet canopy evaporation, and windspeed data can be found at http://climate.geography.hawaii.edu/. Rainfall data is available at https://www.hawaii.edu/. Land use, stream, road, building, and fire station data are found at https://planning.hawaii.gov/.

Additional information

Funding

This work was supported by Innovative Human Resource Development for Local Intellectualization program through the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (IITP-2024-RS-2023-00260267) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2023R1A2C1007742).