Assessing the predictive efficacy of six machine learning algorithms for the susceptibility of Indian forests to fire

Keywords: artificial neural networks, boosted logistic regression, classification and regression trees, forest fire, k-nearest neighbours, machine learning, MODIS, support vector machine, susceptibility mapping.

Fires in tropical and temperate deciduous forests are detrimental, as these forest landscapes are not adapted to regular or intense burning, this therefore adversely impacting their ecological and commercial value (Juárez-Orozco et al. 2017; Harrison et al. 2021). Large-scale and intense fires have not been part of natural disturbances of tropical rainforests. However, their intensity and severity in recent decades (Herawati and Santoso 2011), mainly owing to warm and dry conditions associated with changing climate have significantly increased (Harrison et al. 2021). According to the FAO (2020), forest fires are one of the leading drivers of forest degradation each year. Hansen et al. (2013) reported ~4.2 million km² of gross global forest loss from 2001 to 2018. In 2015 alone, ~98 Mha of forested area was burned, especially in the tropics (FAO 2020). In fact, fire-induced tropical forest loss accounts for 69% of total carbon addition to the atmosphere (Baccini et al. 2017; Armenteras et al. 2021). Large-scale forest loss due to fire hazards can considerably decrease the terrestrial carbon sink, and thereby alter regional weather and global climate at large (Bonan 2008; Swann et al. 2018; van Wees et al. 2021).

There is evidence to show that forest fires may become more severe and intense in the future owing to climate change (USGCRP 2017; Artés et al. 2019; Brown et al. 2021). Jolly et al. (2015) reported that the length of the fire season increased by 18.7% from 1979 to 2013; moreover, 25.3% (29.6 million km²) of the global vegetated surface was exposed to fire seasons. Forest fires are complex; their events and behaviour are determined by complex and non-linear factors, including human activities, weather and climate conditions, vegetation types, and the greater topography (Jain et al. 2020; Zhang et al. 2021). Human activities, for instance, rapidly change the land use of forest landscapes, which in turn increases the vulnerability of forests to fire incidents. Climate change, however, causes longer, more frequent and stronger dry and warmer spells (Mozny et al. 2021; Gannon and Steinberg 2021). It also creates an increased El Niño–Southern Oscillation (ENSO) and sea surface temperature anomalies (Chen et al. 2011; Cai et al. 2017), which are directly associated with increasing fire-prone areas worldwide (Burton et al. 2020). Drought leads to fuel accumulation, and a rise in surface temperature, which increases forests’ flammability (Brando et al. 2019; Ma et al. 2020), thereby triggering frequent and severe fires (De Faria et al. 2017). Additionally, the interactions of climate factors with the natural topography and wind speed direction result in severe and extended forest fire events (Buma 2015; De Faria et al. 2017; Brando et al. 2019; French 2020; Armenteras et al. 2021).

Earth observation satellites provide timely and repetitive information on active and archived fires at regional and global scales (Chuvieco et al. 2019). Remote sensing datasets have widely been used to map active fires and burned areas (Chuvieco et al. 2019). Remote sensing satellite data of MODIS (Moderate Resolution Imaging Spectroradiometer) (Pourtaghi et al. 2016; Chuvieco et al. 2019), Landsat (Syifa et al. 2020) and Sentinel-2 (Navarro et al. 2017; Lang et al. 2019; Roteta et al. 2019) provide information on the extent of fire-affected areas, burn severity and fire susceptibility. Over the last two decades, many researchers have also used MODIS fire hotspot data to integrate it with different predictive modelling approaches for mapping the susceptibility of forests to fires (Ma et al. 2020; Mohajane et al. 2021; Sulova and Joker Arsanjani 2021).

Machine learning (ML) algorithms are popular and emerging predictive approaches in wildfire science and management (Jain et al. 2020). Forest fire susceptibility prediction can be made using ML prediction models (Pourghasemi et al. 2016; de Bem et al. 2018), which define vulnerable areas based on the correlation between wildfire occurrence and sets of predictors. ML algorithms are unique owing to their efficiency, powerful computation, noise handling and ability to capture non-linear and dynamic relationships among variables (Bianco et al. 2019). A review by Jain et al. (2020) found 300 publications on ML applications in wildfire science and management up to 2019. Table 1 shows some of the recent studies that have used different ML algorithms for mapping forest fire susceptibility in some of the other fire-prone areas globally.

Table 1.  List of recent studies on forest fire prediction and mapping using ML algorithms.

Kale et al. (2017) stated that mainly anthropogenic activities initiate forest fires in India. However, climate change and ENSO events create favourable conditions for spreading severe and intense fire. ENSO and Indian monsoon rainfall have an inverse relationship. It is well documented that ENSO events are associated with rainfall deficits, which cause warmer and dry spells (Azad and Rajeevan 2016). According to the Forest Survey of India-published Indian State of Forest Report (ISFR 2021), 35.46 % of forest cover in India is fire-prone. Extremely prone area is 2.81%, very highly prone is 7.85% and highly prone is 11.51%. Further, 45–64% of forests in India may face the effects of climate change by 2030. According to World Bank (2018), out of the 647 districts in India, 380–445 districts had fire events every year from 2003 to 2016. Among the major forest types in India, the dry deciduous broadleaved forests are found to be highly susceptible to forest fires. Sannigrahi et al. (2020) reported highly concentrated MODIS active fire alerts in India’s central and east-central states (Odisha, Chhattisgarh and Madhya Pradesh). These regions are mainly covered by deciduous forests, which are highly susceptible to seasonal forest fires. According to the World Bank Group (2021), climate extremes would be intense in the future in India, with increased drought risk and rainfall uncertainties. Northern India would witness a significant temperature increase, whereby annual minimum and maximum temperatures would be expected to increase to a larger degree than the country’s mean temperatures.

Although numerous wildfire prediction models are available worldwide, they are not yet sufficiently efficient to be adopted at different scales and for different geographical conditions. In India, both burned area and the number of fires show increasing trends. Previously, no study has explored the different ML algorithms and predictors that are used in this study to create fire prediction models. Therefore, we attempted to assess the forest fire susceptibility mapping of Indian forests of six ML algorithms, and compared their predictive efficacy, using artificial neural network (ANN), boosted logistic regression (BLR), classification and regression trees (CART), k-nearest neighbours (KNNs), penalised logistic regression (PLR) and support vector machine (SVM). In addition, we assessed the key driving factors for forest fires in the study region. Twenty years (2001–2020) of annual average data of forest, climatic and topography predictors were analysed and used in ML algorithms for fire susceptibility prediction based on 20 years of MODIS fires hotspots training data. This research suggests the best-suited algorithm for creating a fire prediction ML model, using minimal parameters related to forests, climate and topography. This study should help decision and policymakers with effective forest fire management.

This study was conducted in India, a South Asian and the seventh-largest country in the world. India lies between 8°4′ and 37°6′ N latitude and 68°7′ and 97°25′ E longitude (Fig. 1). The geographical area of India is 2.4% of the world’s land area and it holds 1.7% of the global forest area (FAO 2010; Rajashekar et al. 2018). Currently, India’s forest cover is 713 789 km², constituting 21.71% of its geographical area (ISFR 2021). The mean annual temperature is 24.83°C. India receives 1170 mm of rainfall annually, 80% of it during monsoon months (Praveen et al. 2020). Broadly, India can be categorised into four climatic regions: North-west, North-east, Central and Peninsular India (Champion and Seth 1968; Guhathakurta and Rajeevan 2008). Vegetation varies from the Himalayas in the north to the Western Ghats in the south, moist and dry deciduous in the Central, sparse and thorny in the North-west and wet evergreen forest in the North-east of India (Reddy et al. 2015). India displays nearly all types of climates owing to its physiographic position (Martínez-Austria et al. 2016). The four seasons include winter (January–February), summer (March–May), monsoon (June–September) and post-monsoon (October–December) (Das et al. 2015). Elevation varies from coastal zones to the world’s highest mountain ranges (Reddy et al. 2015).

Fig. 1.  The location of the study area shows different landcover types from International Geosphere-Biosphere Programme (IGBP) MODIS classification data (map was created using ArcGIS Desktop 10.5).

The first step in the methodology includes the data collection and analysis, including MODIS fire hotspots, forest, climatic and topographic data parameters. Fig. 2 shows an year wise variations in the number of fires and burned area. These datasets had an inconsistent spatial resolution (Table 2), so were resampled to a spatial resolution of 1 km (same as MODIS fire hotspot), followed by data filtering. All extracted data were split into train and test with a 70 and 30% ratio, respectively, followed by data normalisation. The normalisation process ranges the data between 0 and 1, which reduces data inconsistencies (Murtaza et al. 2020). A scatterplot matrix was generated to obtain the scatterplots, Pearson correlation and distribution of predictors with fire and non-fire data class. The over–under sampling method was used to remove the class imbalance problem. We trained six ML models, namely ANN, BLR, CART, KNN, PLR and SVM. A 5-fold cross-validation was performed to avoid the model’s overfitting and underfitting problem, followed by hyperparameter tuning. The relative importance was assessed to know the predictor’s importance in fire probability prediction. ML models evaluation was performed using the Area Under the Receiver Operating Characteristics Curve (ROC/AUC) (Bradley 1997). The optimal ML models were selected based on the largest ROC/AUC metrics value, followed by prediction on test data. Finally, forest fire susceptibility maps were generated and analysed for class-wise fire probability (Fig. 3).

Fig. 2.  Variations in burned area and number of fires in India (Data source: Global Wildfire Information System, GWIS – https://gwis.jrc.ec.europa.eu/).

Fig. 3.   Flow chart of the methodological framework followed in this study. (NDVI: normalized difference vegetation index; LC: land cover; RH100: relative height at top of canopy; LST: land surface temperature; Tmax: Maximum temperature; Tmin: minimum temperature; CWD: climate water deficiit; AET: actual evapotranspiration; KBDI: Keetch-Byram drought index; DEM: digital elevation model).

Table 2.  List of data type, parameters used from sensors with spatial resolution, time-period and their download source.

MODIS Terra Collection 6.1 fire hotspots from 2001 to 2020 were downloaded from the Fire Information for Resource Management System (FIRMS) (https://earthdata.nasa.gov/firms). These data files contain the latitude, longitude, acquisition date and time, and confidence (C) (range 0–100%). The confidence class range between 0 and 30% has low, 30–80% has nominal and 80–100% has high fire detectability (Giglio et al. 2018).

Forest fuel type, health and canopy height are the important deciding factors in forest fires. Forest types in the study region are fuel types (dry or moist vegetation or grasslands). Forest type land cover (LC) was used from the MODIS International Geosphere-Biosphere Programme (IGBP) Land Cover Type Product (MCD12Q1), having a spatial resolution of 500 m. Normalized Difference Vegetation Index (NDVI) is an effective indicator of vegetation health status (Huang et al. 2021), which can be correlated with fires. MODIS monthly mean NDVI from 2001 to 2020 was downloaded from Climate Engine (CE) (Huntington et al. 2017). The annual mean composite of NDVI was averaged to obtain a single NDVI raster. NDVI ranges from −1 to 1, which represents vegetation health status. Forest gridded mean canopy relative height (RH100) metrics with a spatial resolution of 1000 m were derived from Global Ecosystem Dynamics Investigation (GEDI) and available from the Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC) (Table 2).

We used the mean values of eight climatic data parameters from 2001 to 2020 (Table 2). These climatic parameters are actual evapotranspiration (AET), climate water deficit (CWD), maximum temperature (Tmax), minimum temperature (Tmin), precipitation (PPT), land surface temperature (LST), wind speed (WS) and Keetch–Byram drought index (KBDI). AET, CWD, Tmax, Tmin, PPT and WS data from TerraClimate (Abatzoglou et al. 2018) and LST data from MODIS were downloaded using CE, whereas KBDI (2007–2020) was obtained from the Google Earth Engine (GEE) (Gorelick et al. 2017) cloud computing platform. KBDI, derived from daily meteorological data, is widely used in wildfire prevention and forecasting. Its scale ranges from 0 to 800, with lower values indicating no moisture deficit and low fire potential and higher values indicating the opposite (Takeuchi et al. 2015).

Topographic data parameters, including digital elevation model (DEM), were obtained from the National Aeronautics and Space Administration (NASA) Shuttle Radar Topography Mission (SRTM) 30 m data using GEE. The slope and aspect map was derived from DEM in ArcGIS Desktop 10.5.

For 20 years (2001–2020), the downloaded fire hotspots were merged, cleaned and filtered. We used eight LC forest classes of MODIS data as a mask to remove fire hotspots from non-forest areas such as croplands, water bodies, barren land and settlements. Further, fire hotspots from January to June were used to avoid the inclusion of crop stubble burn pixels from September to November, and cloud-covered pixels from the monsoon period. Only confidence values greater than 85% were used as high-intensity true fire hotspots or pixels alarms. Finally, we obtained the 20-yeardistribution of 95 999 fire hotspots and pixels shown in Fig. 4.

Fig. 4.  Distribution of 20 years’ (2001–2020) forest fire hotspots or pixels in Indian forests.

Out of 17 MODIS LC classes, 8 LC classes were selected for our analysis, as shown in Fig. 5a. Canopy height (RH100) data are sparsely distributed over the study area. Therefore, the Bayesian Empirical Kriging approach was used to interpolate the missing data values using the Geostatistical Analyst tool in ArcGIS Desktop 10.5. RH100 data was classified into five classes (<10, 10–20, 20–30, 30–40 and >40 m) (Fig. 5b). NDVI range was classified into four classes, namely dead vegetation or objects (<0), unhealthy vegetation (0–0.33), moderately healthy vegetation (0.33–0.66) and very healthy vegetation (0.66–1) (Fig. 5c).

Fig. 5.  Forest predictors used in the current study: (a) landcover (LC); (b) mean relative height (RH100); and (c) NDVI.

All the climatic raster maps were arranged into different classes as shown in Fig. 6a–g. Both AET (Fig. 6a) and CWD (Fig. 6b) were categorised into five classes, namely very low (<300 mm), low (300–600 mm), moderate (600–900 mm), high (900–1200 mm) and very high (>1200 mm). The five categories of KBDI represent no fire risk (<100), low fire risk (100–200), medium fire risk (200–300), high fire risk (300–400) and very high fire risk (400–540) (Fig. 6c). LST was classified into very low (<10°C), low (10–20°C), moderate (20–30°C), high (30–40°C) and very high (>40°C) (Fig. 6d). PPT was classified into very low (<500 mm), low (500–1000 mm), moderate (1000–2000 mm), high (2000–3000 mm) and very high (>3000 mm) (Fig. 6e). Tmax was classified into low (<10°C), moderate (10–20°C), high (20–30°C) and very high (>30°C) (Fig. 6f). Tmin was classified into very low (<5°C), low (5–10°C), moderate (10–15°C), high (15–20°C) and very high (>20°C) (Fig. 6g). WS was classified into very low (<1 m s^–1), low (1–1.5 m s^–1), moderate (1.5–2.0 m s^–1), high (2.0–2.5 m s^–1) and very high (>2.5 m s^–1) (Fig. 6h).

Fig. 6.  Climatic predictors used in the current study: (a) AET; (b) CWD; (c) KBDI; (d) LST; (e) PPT; (f) Tmax; (g) Tmin; and (h) WS.

Topographic parameters maps of aspect, DEM and slope were also categorised into different classes (Fig. 7). Aspect was classified into five classes (<75, 75–50, 150–225, 225–300 and >300°) (Fig. 7a). DEM was categorised into <400, 400–800, 800–1200, 1200–1600 and >1600 m (Fig. 7b), and slope into <2.5, 2.5–5.0, 5.0–7.5, 7.5–10.0 and >10.0% (Fig. 7c).

Fig. 7.  Topographic predictors used in the current study: (a) aspect; (b) DEM; and (c) slope.

We obtained 75 577 MODIS fire hotspots or pixels over the Indian forests. Also, 24 024 non-fire points were randomly generated manually from non-fire point zones of India. We had a total of 99 601 data points containing fire and non-fire points. However, 6453 data points had null values and were excluded. Subsequently, we had 93 148 data points, split into train and test (70:30%). Therefore, 65 204 points were used as training data, and 27 944 points were the test data. Of 65 204 training data points, 49 566 were fire points, and 15 638 were non-fire points. Therefore, there is a class imbalance situation, as there are fewer non-fire points than fire points. We used the Random Over-Sampling Examples (ROSE) library (Lunardon et al. 2014) ‘ovun.sample’ function in RStudio (RStudio Team 2021) on training data to avoid class imbalance. We obtained 32 733 as non-fire points and 32 471 as fire points using the over–under class sampling approach using the ROSE library.

After creating training and testing datasets, six ML algorithms were trained using only training datasets, and test data were used for models performance evaluation for unseen conditions. A 5-fold cross-validation technique was used to avoid overfitting or underfitting problems. We used the caret package (Kuhn 2008) in RStudio for models buildings. From the caret package, we used ‘nnet (ANN)’, ‘LogitBoost (BLR)’, ‘rpart (CART)’, ‘knn (KNN)’, ‘plr (PLR)’ and ‘svmRadial (SVM)’ ML models libraries.

ANN is a mathematical model designed to mimic a biological neural network (Zhang et al. 1998), with neurons as the basic building blocks. Firstly, all input values are multiplied by individual weights. Then, all weighted inputs and biases are summated and then passed through the activation function, also known as the transfer function. Neural networks consist of input, output and hidden layers. The input and output are visible, and the hidden layer is non-observable (Maeda et al. 2009). Training neural networks (nnet) using the caret package needs size and decay hyper-parameters. Size is the number of units in the hidden layer (‘nnet’ fits a single hidden layer), and decay is the regularisation parameter to avoid overfitting.

Eqn 1, show the mathematical expressions for ANN, where x_i is the input, w_i is the weight, f is the activation function and b is the bais.

BLR ‘LogitBoost’ is a boosting classification algorithm that performs an additive logistic regression and minimises the logistic loss (Friedman et al. 2000). LogitBoost was designed as an alternative solution to address the downsides of Adaboost in holding noise and outliers in data (Kamarudin et al. 2017). In LogitBoost, the dependent variable was a binary value representing the presence (1) or absence (0) of forest fires. The tuning parameter in BLR is ‘nIter’ (number of boosting iterations). In BLR, the input data set is N = {(x₁, y₁),……, (x_i, y_i),… (x_n, y_n)}, where x_i ∈ X and y_i ∈ Y = {−1, +1}. The model can be expressed as Eqn 2

sign[F(x)] is a function that has two possible output classes; k is the number of input iterations

CART does not develop a prediction equation. The data are segregated along the predictor axes into subsets with the homogeneous dependent variable (Krzywinksi and Altman 2017). CART involves identifying and constructing a binary decision tree using a training data sample for which the correct classification is known (Breiman et al. 2017). The CART algorithm’s hyperparameter ‘complexity parameter' (cp) determines how deep the tree will grow. Here, it is assigned a small value, which will allow a decision on further pruning. We want a cp value for pruning the tree, minimising the xerror (cross-validation error).

KNN is a supervised machine learning algorithm that uses previously memorised data to classify new data points into the target class depending on the nearest available points (Wu et al. 2018). The performance of KNN in computing the prediction areas depends on D, the distance between similarities; k, the number of nearest neighbours to be used when calculating predictions, and the scheme to weight individual neighbours when computing predictions (Chirici et al. 2016). The model can be expressed by Eqn 4

The PLR generalises the standard logistic regression (LR) with a penalty term on the coefficients (). PLR can be fitted in the regularisation framework with loss + penalty (Wahba 1999; Gao et al. 2000; Park and Liu 2011). The loss function controls the model’s goodness of fit, and the penalisation term helps to minimise overfitting so that generalisation can be obtained. The PLR uses the unbounded logistic loss, making the classifier sensitive to outliers (Park and Liu 2011). Wahba (1999) showed that the linear PLR is equivalent to Eqn 5:

where l is the binomial log-likelihood, and lambda (λ) is a positive constant. As a result of the quadratic penalisation, the norm of the coefficient estimates is smaller than regular LR; however, none of the coefficients is zero. The penalty term measures the smoothness to avoid overfitting, and the tuning parameter λ decides how smooth the PLR model will be. Therefore, the hyperparameter λ choice significantly affects the resulting model (Park and Liu 2011).

SVM is one of the most robust and accurate machine learning techniques, which separate the classes present in the data by identifying the optimal hyperplane of separation (Wu et al. 2008; Rodrigues and De la Riva 2014). SVM was introduced by Vapnik (1995) and followed the statistical learning theory and structural risk minimisation principal methods. Bui et al. (2012), Naghibi et al. (2018) and Jaafari and Pourghasemi (2019) also trained the SVM algorithm to create a hyperplane that segregates the classes into fire and non-fire data. SVM’s tuning parameter ‘cost' (C) defines the possible misclassifications. This hyperparameter imposes a penalty to the SVM model for making an error. The higher the value of C, the fewer chances that the SVM algorithm will misclassify the data point. The SVM radial kernel also requires setting a smoothing hyperparameter ‘sigma’ to give the curvature weight of the decision boundary. Eqn 6 represents the general operation of the SVM model for computing the prediction for a binary-class dataset. Here, is weight vector, is an input vector, b = Intercept and bias term of the hyperplane equation

Hyperparameter tuning gives an optimal value of tuning parameters to build an accurate model from a dataset. A tunelength of 10 was used for each model’s hyperparameter tuning. We tuned ‘size and decay for ANN’, ‘nIter for BLR’, ‘cp for CART’, ‘k for KNN’, ‘λ for PLR’ and ‘Cost (C) and sigma for SVM’. Further, pre-processing steps, including centre and scaling, were applied for data normalisation. We used classProbs = TRUE in ‘trainControl’ function of the caret package for predicted outcomes as a probability (0–1).

Most ML models using binary variables are evaluated using ROC/AUC (Grau et al. 2015), which measures the model’s performance. The graph of sensitivity and specificity provides a visual and statistical extent of ML algorithms prediction accuracy (Kalantar et al. 2020). ROC/AUC can be used to validate the prediction of various ML models against an original training dataset. Therefore, the ROC/AUC technique is a productive method for portraying the efficiency of a probability map predicted by a particular ML model (Satir et al. 2016). Here, true positive (TP) and true negative (TN) are sample datasets that are correctly classified, and false positive (FP) and false negative (FN) are samples that are misclassified (Bui et al. 2017). The ROC/AUC calculated using the sensitivity (TP rate) (7) and specificity (FN rate) (8) quantifies the performance of the models. ROC/AUC values below 0.6 indicate poor performance, values of 0.7–0.8 denote good performance, values of 0.80–0.90 represent very good performance and greater than 0.90 shows an excellent performance of the model (Bui et al. 2017; Jaafari and Pourghasemi 2019).

After evaluating the model’s performance and desired accuracy using the ROC/AUC method, forest fire susceptibility mapping was performed on a raster stack over the study region. Finally, the computed probability was displayed as a map, which shows the forest fire susceptible zones, and parameters class-wise mean fire probability was calculated using the zonal statistics tool in ArcGIS Desktop 10.5.

Fig. 8a–h shows the variations in the annual mean of climatic parameters and the number of fires. Fig. 8a shows that the number of fires was at a maximum in 2012 when AET was 823.37 mm. In high AET years, such as 2006, 2010, 2011, 2013, 2014 and 2015, fires decreased. The number of fires and CWD in 2009, 2016 and 2016 was higher than in other years. In 2015, CWD was lowest at 391.54 mm, and the number of fires sharply declined (Fig. 8b). Drought index KBDI had a higher value in 2009 and 2012 as did the number of fires, while in 2019, the KBDI was less than 200, and the number of fires also declined. In 2015, the reverse trend was observed as the KBDI was high; however, the number of fires sharply declined (Fig. 8c). The number of fires also shows high peaks in 2009, 2012 and 2018 when LST was 26.04, 25.35 and 25.06°C respectively (Fig. 8d). Maximum fire incidents were observed in 2012 when NDVI was low at 0.543. In 2013, 2014 and 2015, NDVI showed increasing trends, decreasing the number of fires. The maximum NDVI of 0.581 was in 2015, and the number of fires was also lowest between 2009 and 2019.

Fig. 8.  Variations in annual mean climatic parameters and the number of fires: (a) AET; (b) CWD; (c) KBDI; (d) LST; (e) NDVI; (f) PPT; (g) Tmin; (h) Tmax.

Further, in 2019, the NDVI was high at 0.573, and the number of fires declined (Fig. 8e). Fig. 8f shows that fires rose with PPT decline in 2009, 2012 and 2018. Also, the years 2010, 2011, 2013 and 2019 recorded higher PPT and a lower number of fires. In 2015, the number of fires was the lowest between 2009 and 2019 as this year received precipitation of >1700 mm annually (Fig. 8f). Likewise, between 2009 and 2019, fire incidents peaked in 2009, 2012, 2016 and 2018 when Tmax was 27.27, 26.41, 26.98 and 26.34°C respectively. The lowest number of fire incidents was observed in 2011, 2015 and 2019 when Tmax was 26.51, 26.42 and 26.86°C respectively (Fig. 8h). As Tmax sharply rose from 2015 (26.42°C) to 2016 (26.98°C), the number of fires from 2015 to 2016 also sharply increased (Fig. 8h).

The Pearson correlation (r) among parameters during the fire and non-fire events was computed and plotted into a scatterplot matrix structure, as shown in Fig. 9. A high correlation between LST and KBDI (r = 0.766) was observed, and both of these parameters were inversely related to NDVI (r = −0.339, −0.291). Also, a correlation (r = 0.413) between PPT and NDVI was observed for fire class; however, they showed an inverse relation with CWD (r = −0.458, −0.698). Tmax and NDVI showed an inverse correlation (r = −0.291) with fire class. Tmax and LST strongly correlate (r = 0.849) with fire class. Also, RH100 and slope are positively correlated (r = 0.382) with fire class. AET and NDVI correlate (r = 0.687) with fire class.

Fig. 9.  Scatterplot matrix shows the scatterplots and Pearson correlation among the predictor parameters and training binary class fire and non-fire.

Fig. 10 shows the variations in ROC/AUC across tuning parameters of the ML models, and Fig. 11 compares ROC/AUC, sensitivity (Sens) and specificity (Spec) of the optimal ML models. The optimal model was selected based on the largest ROC/AUC value. ROC/AUC variations across tuning parameters show that the optimal ANN model has a final ROC/AUC value of 0.903 at size = 19 (Fig. 10a), while sensitivity and specificity were 0.860 and 0.778 respectively. The final ROC/AUC value of 0.838 was used for the optimal BLR model at nIter = 81(Fig. 10b), while sensitivity and specificity were 0.849 and 0.710 respectively. For optimal CART, the final value ROC/AUC was 0.842 at cp = 0.001817006 (Fig. 10c), while sensitivity and specificity were 0.849 and 0.756 respectively. The optimal KNN was obtained at k = 5, for which ROC/AUC is 0.937 (Fig. 10d), while sensitivity and specificity were 0.842 and 0.867 respectively. The final values ROC/AUC of 0.839 used for the PLR model were λ = 0.001333521 (Fig. 10e), while sensitivity and specificity were 0.852 and 0.708 respectively. For SVM training, the tuning parameter sigma was held constant at 0.1278743. The final ROC/AUC value of 0.927 was used for the SVM model at σ = 0.1278743 and C = 128 (Fig. 10f), while sensitivity and specificity were 0.907 and 0.813, respectively (Fig. 11).

Fig. 10.  Variations in ROC/AUC across tuning hyperparameters of ML models: (a) ANN; (b) BLR; (c) CART; (d) KNN; (e) PLR; and (f) SVM.

Fig. 11.  Box plots to compare ROC/AUC, sensitivity (Sens) and specificity (Spec) of the optimal ML models.

Fig. 12a–f shows the overall importance (scaled from 0 to 100) of the predictors in the ML models for fire prediction. All climatic parameters such as Tmax, CWD, LST, Tmin, PPT and AET have a high importance score (>50) for fire probability prediction by the ANN model. Forest parameter RH100 is the least important predictor (Fig. 12a). NDVI is the most important predictor in the CART model, whereas climatic parameters WS, Tmax and CWD have a high overall importance score in the CART model. The slope is the least important predictor in the CART model (Fig. 12c). The KNN, BLR, PLR and SVM models have the same importance for their respective predictors in fire probability prediction. Fig. 12b, d, e, f shows that forest parameter NDVI is the most important predictor, and LC and RH100 have a high overall importance score for fire prediction.

Fig. 12.  Overall importance of predictors in the ML models: (a) ANN; (b) BLR; (c) CART; (d) KNN; (e) PLR; and (f) SVM.

The ROC/AUC curves in Fig. 13a compare the ML models for prediction rate, while ROC/AUC curves in Fig. 13b compare the ML models for success rate. Fig. 13a shows that SVM and ANN are the best-performing models and have the highest prediction rate with ROC/AUC of 0.908 and 903 respectively, and BLR has an ROC/AUC of 0.802, which is the lowest (Table 3). The success rate ROC/AUC curve shows that KNN is the best-performing (Fig. 13b) with the highest ROC/AUC of 0.945, SVM has 0.930, and ANN has 0.904, whereas BLR has 0.808, which is the lowest.

Fig. 13.  Prediction rate and success rate curve for the ML models.

Table 3.  Prediction and success rate ROC/AUC of the ML models.

Fig. 14a–f shows the forest fire susceptibility maps of the Indian forest regions predicted using the six ML models. Models slightly vary in their outcomes of fire probability prediction maps in different parts of India; however, all six ML models predicted a high to a very high probability of forest fire in the North-east regions of India. The Central Indian forest regions are also susceptible, as the probability of forest fire is moderate to high. The Himalayan forests in Uttarakhand and Himachal Pradesh states (parts of the lower Himalayas) have moderate to very high susceptibility to fires. The Western Ghats of India have low to moderate susceptibility, while the Eastern Ghats and Peninsular Indian forest regions have moderate to high susceptibility to forest fires. North-western parts of India have very low to low susceptibility. Coastal zones and the upper Himalayan region of India have very low to low forest fires susceptibility.

Fig. 14.  Forest fire susceptibility mapping over the Indian forest regions: (a) ANN; (b) BLR; (c) CART; (d) KNN; (e) PLR; and (f) SVM.

Table 4 shows the class-wise mean of forest fire probability predictors in the ML models. Models predicted that Evergreen broadleaf forests have maximum fire probability (0.55–0.64) among other LC classes. The best-performing SVM model predicts that precipitation class 2000–3000 mm has a maximum mean fire probability of 0.44. Tmax in the range of 20–30°C has a maximum mean fire probability of 0.45. NDVI with moderately healthy vegetation (0.33–0.66) and very healthy vegetation (0.66–1) has a fire probability of 0.25 and 0.54, respectively. RH100 in the range of 20–30 m has a maximum mean fire probability of 0.44, while RH100 >40 m has the lowest mean fire probability of 0.11. KBDI in the range of 200–300 has the maximum mean fire probability of 0.43.

Table 4.  Mean of the class-wise forest fire probability of predictors in the ML models.

As in several parts of the world, forest fires and burned area have dramatically increased in Indian forests. Although numerous wildfire prediction models are available worldwide, they are not yet efficient enough to be adopted on a large scale and in different geographical conditions. Our findings suggest the best-suited algorithm for creating a fire prediction ML model, using minimal important parameters related to forests, climate and topography. We noted that climate parameters, such as PPT, CWD, Tmax, LST and KBDI are associated with the number of forest fires. Specifically, we noted that the forest parameter NDVI has an inverse link to number of fires; for instance, when NDVI increases, the incidence of forest fires decreases; this was specifically observed in 2011, 2015 and 2019. Additionally, it can be seen that all forests, climatic and topographic parameters have a complex and mutual relationship in increasing the intensity and number of forest fires. Therefore, efficient models that determine complex, non-linear and dynamic relationships between parameters are helpful for accurate fire susceptibility prediction.

Of the six ML models that we used, SVM and ANN models outperformed the others, and yielded an excellent performance (>0.90) in terms of the prediction rate. The other four models, namely KNN, PLR, CART and BLR, had an ROC/AUC in the range of 0.80–0.87. The high predictive efficacy of SVM is possibly due to its ability to deal with dynamic relationships, handling complexity in data, and be least affected by noisy data, and less prone to overfitting (Ballabio and Sterlacchini 2012; Pham et al. 2016). It may also be noted that SVM is an advantageous algorithm for binary classification problems, such as non-fire and fire class. Similarly, the excellent performance of ANN was due to its capacity to recognise hidden relationships between complex and non-linear datasets. The performance obtained from SVM and ANN also supports extant literature, which showed similar performance of the SVM model in different fields, including flooding and landslide susceptibility mapping (Sakr et al. 2010; Li et al. 2011; Ballabio and Sterlacchini 2012; Pham et al. 2016; Bui et al. 2017; Wu et al. 2018). ANN showed similar predictive performance to that observed in the current study in various wildfire modelling studies (Lee et al. 2012a, 2012b; Satir et al. 2016). De Vasconcelos et al. (2001), Bisquert et al. (2012), Jafari Goldarag et al. (2016), Adab (2017) also obtained an improved performance of the ANN model in binary classification. Kumar and Kumar (2020) generated fire detection and classification models using MLP and KNN algorithms and compared them, and their findings showed that the MLP algorithm had a higher accuracy (99.96%) than the KNN algorithm. Even if KNN had a better success rate ROC/AUC value (0.945), it was considered biased, and could fail, especially if there were no nearest values (Magnussen et al. 2010). Compared with both SVM and ANN, the KNN, BLR, PLR and CART models have been slightly less accurate in our study. However, these require only one hyperparameter tuning, and less time in training. The advantage of ML techniques over traditional methods is that the former (i.e. ML techniques) can be handy with noisy data, and thereby overcome uncertainties even with limited observations. However, to increase the accuracy of the ML models, the quality of the input data in ML models is also important to consider (Pourghasemi et al. 2020).

Analysis of parameters shows a mutual, complex and interconnected relationship between the predictors and the number of fires. The climatic parameters Tmax, CWD, LST, Tmin, PPT and AET had high overall importance for fire prediction by the ANN model. Generally, these parameters also have an inverse correlation with forest parameter NDVI. Notably, NDVI is the most important parameter for fire prediction in SVM, BLR, KNN, PLR and CART, while forest parameters such as LC and RH100 also have high importance. Previous studies found that these climatic and forest parameters are important for predicting forest fire susceptibility. For instance, Bui et al. (2019) found that NDVI is the main factor that influences forest fire mapping. Pourtaghi et al. (2015) also found the most important factors were NDVI, land use, soil and annual temperature for forest fire susceptibility mapping in the Minudasht forests of Iran. WS is the second-most important predictor after NDVI in SVM, BLR, KNN, PLR and CART. Achu et al. (2021) found that WS is important in forest fire prediction, as our study suggested. Williams et al. (2019) revealed that wind events and delayed onset of winter precipitation are the dominant wildfire triggers.

Although some parameters were considered in the context of their effects and forest fire susceptibility mapping, all the algorithms used effectively operated efficiently in the study region. Satir et al. (2016) considered the role of anthropogenic factors insignificant in fire susceptibility prediction. However, Achu et al. (2021) demonstrated that anthropogenic factors, such as land use and distance to roads, are important in forest fire modelling. In addition, some studies found that human-related variables have a stronger influence than climate-related variables for fire ignition (Pham et al. 2020; Mohajane et al. 2021). However, in tropical countries like India, where forests are generally non-adapted to fire and climate change has a large impact on forests, the role of climate-related variables in forest fires is very important to investigate. Human activities largely act as fire ignition sources; however, fire intensity and severity depend on the climatic, forest, topographic and weather conditions of the area. Initiation of forest fires in India is generally anthropogenic; however, intensity and severity largely depend on forests, climatic and topographic factors. The present study ignored factors like distance to roads, rivers and settlements, because data on these factors are hard to obtain at a local scale, especially for a country-level study. However, we did consider factors such as NDVI, forest canopy height and land cover that are directly or indirectly impacted by anthropogenic activities.

Forest fire susceptibility maps show that large areas of North-east India’s forests have a very high susceptibility to forest fires. The North-east regions of India do face a huge human burden from activities like shifting cultivation and deforestation. The lower Himalayan region, Eastern Ghats, and Peninsular regions also are highly susceptible to forest fires, specifically because these regions are vulnerable to climate change and experience large variations in climatic factors. The Western Ghats, however, have low to moderate susceptibility, while coastal zones and the upper Himalayan region have very low to low susceptibility. Additionally, North-western India also has a very low to low fire probability due to the presence of sparsely vegetated areas. Our fire susceptibility maps also support FSI ISFR reports on the biannual Indian forest cover assessment, indicating that these regions are prone to forest fires. Therefore, our prediction models based on MODIS data and two decades of mean climatic parameters represent accurate outcomes. It should also be noted that many researchers in the past have stated that climate change will also have an impact in terms of the intensity of wildfires in the future, as evidenced by increased mean temperature, drying conditions and weather patterns. Based on these findings, and observations from extant literature, this study does support climate change linked to wildfires and forest fires in India, and expresses its apprehensions for the future, when conditions are expected to be more severe.

Dutta et al. (2016), Mayr et al. (2018), Jain et al. (2020) suggested that the ensemble ML approaches perform better than a single classifier. Ensemble ML models like RF and boosting algorithms (XGBoost) have offered improved accuracy in some previous studies (Milanović et al. 2021; Mohajane et al. 2021); however, we believe that these ensemble models take more time to run than our ML models. In fact, recently, some studies have also highlighted the improved results of deep learning over other methods. For instance, Zhao et al. (2018) and Zhang et al. (2019, 2021) found that CNN was better in classification than SVM. However, more testing on model comparisons, situations criteria and well-established accuracy measurements are needed in order to take care of overfitting and underfitting, using suitable validation approaches.

The current study is possibly among the first attempts to map the forest fire susceptibility of Indian forests, using six ML algorithms and comparing them. This study suggests that SVM and ANN are the best-suited ML algorithms for creating a model to assess the probability of the area being susceptible to fire using minimal parameters. Importantly, both these models showed an excellent accuracy (AUC/ROC > 0.9) in the prediction rate. Further, this study also identified the most to least important forest, climatic and topographic parameters useful in forest fire susceptibility mapping. Climate parameters are uncertain owing to the changing climate; however, they are strongly linked to increased fires and burned area. Forest fire susceptibility maps are helpful in identifying the most to least fire-prone forest regions in India. The ensemble ML and deep learning models possibly improve accuracy; however, they are complex, and training takes considerable computational time. Therefore, we recommend that the inclusion of other fire-related factors and anthropogenic parameters in the best-performing algorithms (SVM and ANN) could improve the accuracy of outcomes. Moreover, seasonal variations in ENSO-related parameters, such as sea surface temperature anomalies, do play an important role in triggering forest fires, and must be identified and used in the modelling process. Furthermore, using ML algorithms for fire susceptibility, prediction should be enhanced, using all available information and data related to fire events before implementation, possibly by using cloud computing platforms like GEE and CE.

The data supporting this study will be shared on reasonable request to the corresponding author.

The authors declare no conflict of interest.

This research did not receive any specific funding.

Laxmi Kant Sharma: conceptualisation, methodology design, formal analysis, reviewing, writing and editing. Rajit Gupta: data curation, methodology design, coding, model run, formal analysis, reviewing writing, visualisation. Naureen Fatima: data curation, formal analysis, writing.