Burned Area Classification Based on Extreme Learning Machine and Sentinel-2 Images

Abstract

Sentinel-2 satellite images allow high separability for mapping burned and unburned areas. This problem has been extensively addressed using machine-learning algorithms. However, these need a suitable dataset and entail considerable training time. Recently, extreme learning machines (ELM) have presented high precision in classification and regression problems but with low computational cost. This paper proposes evaluating ELM to map burned areas and compare them with other machine-learning algorithms broadly used. Several indices, metrics and training times were used to assess the performance of the algorithms. Considering the average of datasets, the best performance was obtained by random forest (DICE = 0.93; omission and commission = 0.08) and ELM (DICE = 0.90; omission and commission = 0.07). The training time for the best model was from ELM (1.45 s) and logistic regression (1.85 s). According to results, ELM was the best burned-area classification algorithm, considering precision and training time, evidencing great potential to map burned areas at global scales with medium-high spatial resolution images. This information is essential to fire-risk systems and burned-area records used to design prevention and fire-combat strategies, and it provides valuable knowledge on the effect of fires on the landscape and atmosphere.

1. Introduction

Wildfires are global phenomena that cause environmental and social transformations, including biodiversity loss, soil degradation, harm to human beings and economic losses [1]. Nowadays, there is growing concern over this matter since, under current climate change conditions, there will be an increase in the season duration, magnitude and consequences of wildfires [2]. In this context, it is increasingly necessary to have precise and detailed information regarding the affected area in order to both assess and consider damages and to implement prevention plans and processes of recovery of the infrastructure and ecosystems affected [3].

Satellite images provided by optical sensors can be used to delimit burned areas resulting from wildfires. This is possible due to the strong decrease of reflectance in the near-infrared (NIR) and the increase in the shortwave infrared spectrum (SWIR) of vegetation as a result of dryness caused by fire [4]. In addition, the difference in the information provided by satellite data can be used to estimate spectral indices, which help to describe the vegetation condition, the severity of the damage or the levels of carbonation of the areas affected by fires. The most common indices are burned-area index (BAI), normalized burn ratio (NBR) and normalized difference vegetation index (NDVI) [5].

In general terms, there are two approaches for the mapping of burned areas using satellite data. The first one is based on rules that identify variations in the burned spectral signals with respect to the unburned environment, based on which fixed or dynamic thresholds are defined for the spectral bands or indices [6]. The second approach is based on machine-learning algorithms that can learn the spectral features of a sample of pixels with labels and recognize those patterns in other areas of the image [7]. Machine-learning (ML) algorithms build a non-linear transformation whose parameters are estimated in a supervised or unsupervised way to solve classification, regression and clustering problems by formulating an optimization problem [8]. Recently, these types of algorithms have been used for mapping burned areas, showing good results with random forest (RF) [7,9], support vector machine (SVM) [7,10], logistic regression (LR) [11] and multilayer perceptron (MLP) supervised learning Neural Networks [7,12].

Despite the advances in the methodologies and algorithms employed, the estimation of burned areas using satellite images depends heavily on the spatial resolution of the sensor used. This demonstrates the significant underestimation of the calculations performed with low-resolution sensors when compared to estimations of higher resolution [13]. Therefore, it is deemed necessary to explore the use of higher spatial resolution platforms and sensors, such as Sentinel-2, in order to deal with these problems.

Moreover, most of the machine-learning algorithms that are frequently employed in image classification increase their computational cost when the size of the data sets increases, leading to longer training and classification times [14]. However, extreme learning machines (ELM) have recently attracted increasing interest in image classification due to their high accuracy levels, which are comparable to the traditional SVM and MLP, but with a training algorithm with reduced computational cost [15].

Considering the ELM algorithm training efficiency and its good generalization, we hypothesized that ELM neural networks can be used to classify burned areas on medium-spatial resolution images, showing a great potential to deal with this problem at scales that require a massive volume of data. Consequently, this work aims to propose the evaluation of mapping burned areas on Sentinel-2 images through an ELM neural network supervised classification, comparing ELM performance in terms of accuracy and training time to several machine-learning algorithms.

The rest of the article is organized as follows: Section 2 presents the methods used to evaluate the ELM performance with other algorithms. Section 3 presents the methodology of the experiments. Section 4 shows the results and their discussion. Finally, Section 5 presents the conclusions.

2. Fundamentals of the Classification Algorithms Applied

In this research, several classification algorithms were used, whose fundamentals are described below.

2.1. Extreme Learning Machine

Extreme learning machine (ELM) neural networks are known for their extremely fast learning speed, showing high levels of performance, compared to the ones found for multilayer perceptron network (MPL) and support vector machine (SVM) [16]. ELM has shown better performance than neural networks with backpropagation and other classification models, in terms of computational efficiency and proper generalization when applied to Landsat satellite images [17].

The ELM algorithm was originally developed for single-hidden layer feedforward neural networks. One of its main features is its simple training since the weights and biases of the hidden layer are randomly created and the output weights can be determined by the resolution of an overdetermined linear system by the use of the Moore–Penrose generalized inverse [16]. The output expression of a single-hidden layer ELM network is as follows:

$$f_L\left(x\right)=\sum _{i=1}^L{\beta }_{i}g\left(w_1\ast x_1+b_1\right),j=1,\dots ..,N$$

(1)

where:

L: the number of hidden neurons.

N: the number of training samples.

${\beta }_i$: output layer’s weight vector.

w: hidden layer’s weight vector.

g: activation function.

b: biases parameters vector.

x: input vector.

Equation (1) can be written in the matricial form:

$$T=H\beta$$

(2)

where H is the hidden layer output matrix and $\beta$ is the weight vector of the output layer. Specifically, the H matrix has the following structure:

$$H={\left[\begin{array}{ccc}g\left(w_1\ast x_1+b_1\right)& ‥& g\left(w_L\ast x_1+b_L\right)\\ ⁚& ‥& ⁚\\ g\left(w_1\ast x_N+b_1\right)& ‥& g\left(w_L\ast x_N+b_L\right)\end{array}\right]}_{NxL}$$

(3)

where:

m: the number of outputs.

H: the hidden layer output matrix of the neural network.

T: the training data labels matrix.

Unlike gradient descent-based neural networks, for ELM, weights of the hidden layer need not be fitted since they are randomly created. In order to train an ELM network, it is necessary to find a $\widehat{\beta }$ least-squares solution of the $H\beta =T$ linear system, whose expression is as follows:

$$\widehat{\beta }=H^{†}T$$

(4)

where $H^{†}$ is the Moore–Penrose generalized inverse of matrix H.

Given a $\aleph =\left\{\left(x_i,t_i\right)|x_iϵ{\mathbb{R}}^n,t_iϵ{\mathbb{R}}^L,i=1,..,N\right\}$ training set, $g\left(x\right)$ activation function, and L number of hidden neurons, the ELM training algorithm has the following steps:

• Step 1: assign arbitrary input weight $w_i$ and bias $bi,i=1,\dots ,L.$.
• Step 2: calculate the hidden layer output matrix H.
• Step 3: calculate the output weight $\beta$ for $\beta =H^{†}T$.

2.2. Multilayer Perceptron

Multilayer perceptron neural networks (MLP) are widely used for satellite images classification [18] and for burned area mapping applications [19]. These networks have an input layer, one or more hidden layers and an output layer [20]. They are fitted using the back-propagation algorithm, which consists of minimizing the output mean squared error of the network using a gradient descent algorithm [19,21]. In general, MLPs have poor performance when the optimization algorithm falls into local minimum and due to the overfitting phenomenon [18]. However, it is possible to improve the optimization algorithm performance through heuristic search techniques [21].

2.3. Random Forest

It is an ensemble model of several decision trees randomly organized and individually trained. Each tree makes a classification decision where the class with the maximum number of votes is determined for the input data [8]. Each tree is independently organized so successive trees are independent from the previous ones [18]. Random forest (RF) has shown good performance in linear and nonlinear models known for balancing bias and variance [7]. This algorithm is widely used on satellite images data due to the high accuracy of its classifications [22]. One of its outstanding features is that it can successfully handle high data dimensionality and multicolinearity since both are insensitive to overfitting [23] and training algorithm is fast [24]. Moreover, RF has high accuracy for datasets external to those considered during training, which shows that spatial autocorrelation has a low impact on the prediction performance [25].

2.4. Logistic Regression

Logistic regression (LR) is a statistical model that can be used to describe the relationship between a dichotomous dependent variable and a series of independent variables [11]. LR has the function of predicting the result of a categorical variable in relation to the predictor variables [18]. It is an efficient tool for burned area mapping since it offers the chance to obtain the probability of a pixel to be classified as burned or unburned [9,26]. LR provides a good probabilistic framework for the development of burned area algorithms since the models obtained are consistent with variations in environmental variables [21].

2.5. Support Vector Machine

Support vector machine (SVM) is a statistical learning algorithm, robust to data noise and adapted to classification and nonlinear regression problems [10]. SVM transforms a nonlinear regression model into a linear model using kernel functions to map the original input space into a new high-dimensional features space [7]. In this high-dimensional space (hyperplane), SVM finds unique solutions to classification and regression problems [18]. The advantage of SVM over traditional classifiers is that it solves learning problems better when only a small number of training samples are available [27]. SVM has shown good classification results for satellite images, presenting low omission errors [19].

3. Materials and Methods

3.1. Hardware and Software Used

Experiments were conducted in a computer with a processor Intel(R) Core(TM) i7-8565U CPU @1.80GHz 2.00GHz, with 16 GB RAM memory DDR4-2400. As regards the software used, the computer had Ubuntu 18 operating system installed. Codes were programmed with Phyton 3.6 with numpy, pandas, gdal, networkx, tensorflow and pylab libraries. In addition, Qgis 3 [28] and Google Earth (https://www.google.com/earth/versions/, accessed on 30 October 2021) were used to visualize images and to generate the training datasets.

3.2. Study Area

The study area corresponds to the territory that extends from the O’Higgins regions to Bío-Bío (33.8º–38.4º S) in Chile (Figure 1). This area has a long history of wildfire occurrence [29], explained by climatic conditions, economic activities and population concentration [30]. Said area presents a semiarid Mediterranean climate, with most rainfalls occurring during the winter period, reaching 350–1300 mm per year, causing a long dry season with low relative humidity, high temperatures and very low precipitations [31]. Since 2010, due to global climatic change, this area has been suffering a prolonged and severe mega drought [32]. The area’s natural vegetation is dominated by Mediterranean shrublands and sclerophyllous forests [33]. Moreover, it presents a high presence of land cover areas created by human action, including forest plantations, agricultural farms and wildland-urban interfaces, where human activities make them likely to suffer from wildfires. The period chosen for this study is the 2016–2017 season, during which the phenomenon of mega wildfires in Chile took place. Their magnitudes and intensities forced the creation of an additional level in the scale of measurement used for the classification of wildfires [34]. The damages recorded affected 596 thousand hectares among O’Higgins and Bío-Bío. The fire destroyed forests, plantations, agricultural farms, meadows and other vegetation masses, affecting more than 2500 homes [35].

3.3. Training Data Extraction and Validation

In order to create the training datasets for the classification algorithms, 20 Sentinel-2 MSI images [36] with cloud cover lower than 30% were considered before and after the fires of the summer of 2017, which mainly occurred between 30 January and 14 February (

Table 1. Sentinel-2 images ID Tiles, cover region, date and time of sensing.

Sentinel-2 Tiling Grid ID	Cover Region	Date	Time
19HBC	O’Higgins	19 January 2017	14:37:31 UTC
19HBC	O’Higgins	18 February 2017	14:37:51 UTC
19HBB	O’Higgins	19 January 2017	14:37:31 UTC
19HBB	O’Higgins	18 February 2017	14:37:51 UTC
19HCB	O’Higgins	19 January 2017	14:37:31 UTC
19HCB	O’Higgins	15 February 2017	14:28:51 UTC
18HYF	Maule	19 January 2017	14:37:31 UTC
18HYF	Maule	18 February 2017	14:37:51 UTC
19HBA	Maule	19 January 2017	14:37:31 UTC
19HBA	Maule	28 February 2017	14:37:51 UTC
18HYE	Ñuble	19 January 2017	14:37:31 UTC
18HYE	Ñuble	18 February 2017	14:37:51 UTC
19HBV	Ñuble	6 January 2017	14:27:41 UTC
19HBV	Ñuble	15 February 2017	14:28:51 UTC
18HXE	Bío-Bío	2 January 2017	14:47:22 UTC
18HXE	Bío-Bío	21 February 2017	14:47:31 UTC
18HXD	Bío-Bío	2 January 2017	14:47:22 UTC
18HXD	Bío-Bío	21 February 2017	14:47:31 UTC
19HBU	Bío-Bío	6 January 2017	14:27:41 UTC
19HBU	Bío-Bío	15 February 2017	14:28:51 UTC

). The MSI sensor has 13 spectral bands with frequencies spanning from the Visible- and Near-Infrared (VNIR) to the short-wave infrared (SWIR). It has a spatial resolution of 10 m, 20 m and 60 m and 5 days temporal resolution [37]. Data were downloaded from Copernicus open-access hub platform (https://scihub.copernicus.eu/, accessed on 18 August 2021) and atmospherically and topographically corrected to obtain surface reflectance (20 m, L2A level) using Sen2cor algorithm [38].

The creation of the dataset was developed by a human expert using false color composite over Sentinel-2 images (Figure 2A), photo interpretation criteria and spectrum analysis to distinguish between burned and unburned areas (Figure 2B).

Based on the image analysis, two datasets were created:

• POST: pixels values from a post-fire image.
• DELTA: pixels values composed by the difference between a pre-fire and a post-fire image.

For the POST dataset, 100 burned and unburned polygons were built, representing 5.98% of the study area. In the other case, for the DELTA dataset, 80 burned and unburned polygons were selected, adding up to near 2.30% of the study area. In order to ensure the consistency of the datasets, the training polygons were overlapped with several spectral indices that have shown the best response in burned area detection [39,40] (

Table 2. Spectral indices used for burned area identification.

Index	Formulation
BAI [41]	${\displaystyle \frac{1}{{(0.1-b4)}^2+{(0.06-b8A)}^2}}$
BAIS2 [5]	$\left(1-\sqrt{\frac{b6·b7·b8A}{b4}}\right)·\left(\frac{b12-b8A}{\sqrt{b12+b8A}}+1\right)$
MIRBI [42]	$(10·b12-9.8·b11+2)$
NBR2 [43]	${\displaystyle \frac{(b11-b12)}{(b11+b12)}}$

where: b4: Reflectance in the red band (centered at 0.665 µm). b6: Reflectance in the near-infrared band 1 (centered at 0.665 µm). b8A: Reflectance in the near-infrared band 2 (centered at 0.865 µm). b11:Reflectance in the shortwave infrared band 1 (centered at 1.610 µm). b12: Reflectance in the shortwave infrared band 1 (centered at 2.190 µm).

).

3.4. Analysis and Filtering of Training Datasets

One of the main steps to achieve a successfully supervised classification is to have a good training dataset [44]. For that, we analyzed the POST and DELTA datasets in order to remove those pixels that could eventually be misclassified by the expert or showed low separability values. We evaluated the mean; median; standard deviation; 1, 2 and 3 quartiles; and minimum and maximum values for the burned and unburned categories, and eliminated the atypical found values. Regarding separability, several indexes or distances can be applied to image classification [45]. However, for burned-area classification, the separability index (SI) [46] (equation (5)) has been broadly used on different sensors and ecosystems [5,9,47] to select spectral bands and indices that provide a higher discriminating capacity (SI $>=$ 0.75) to the classification process.

$$SI=\frac{\left|{\mu }_q-{\mu }_{nq}\right|}{{\delta }_q+{\delta }_{nq}}$$

(5)

where:

${\mu }_q$: mean of the burned category.

${\mu }_{nq}$: mean of the unburned category.

${\delta }_q$: standard deviation of the burned category.

${\delta }_{nq}$: standard deviation of the unburned category.

Training datasets were randomly balanced in order to obtain the same number of pixels for the burned and unburned categories. This allows for the the proper training of the supervised classifiers [22].

3.5. Experiments

The original set was randomly divided into three sets with the following percentages of the total amount of data: 50% for the training set, 20% for the validation set and 30% for the test set. The training set is used to calculate the classifiers parameters, the validation set offers the chance to choose the best model among those generated by the combination of hyperparameters and the test set is used to carry out the objective performance evaluation of each classification algorithm. The following variations to the hyperparameters of the classification algorithms were conducted:

• ELM: the best model was selected among ELM networks with 1, 2, 3, 4,....., 498, 499 and 500 neurons in the hidden layer.
• Random forest: training and tests were conducted with 5, 10, 15, 20,..., 985, 990, 995 and 1000 random trees.
• Logistic regression: training and tests were conducted with 5, 10, 15, 20,..., 985, 990, 995 and 1000 as the maximum number of iterations for convergence.
• Support vector machine: a polynomial core was used that shows a better performance than radial basis kernel [48] and that has been widely applied in this type of classification [49].
• Multilayer perceptron: training and tests were carried out with a single hidden layer multilayer perceptron, from 1 to 100 neurons, with rectified linear unit activation function (relu) and using the Adam stochastic optimizer [50].

In order to choose the best model among the variations generated by the hyperparameters, the average of global accuracy and the DICE coefficients [9] was calculated over the validation set (Figure 3).

According to Padilla et al. [51], obtaining an accurate analysis of burned classifications or products is much more relevant to consider the accuracy of the “burned” category than the “unburned” because the first is more related to the impacts of biomass burning and its atmospheric effects. For that, the performance of the best models for the burned category was evaluated in the test set with DICE coefficient, omission and commission errors, which have been used in several studies that compared accuracies for burned area classifications [9,13,52]. The DICE coefficient measures accuracy with a probabilistic meaning [40], and the accuracy, omission and commission errors are getting from the confusion matrix [53]. In addition, we evaluated the models according to their training time [14]. Finally, the performance of the selected model was compared using an independent validation area. The validation area was built from a supervised classification of Sentinel-2 (20 m) images from a section of 247,401 hectares from the Maule Region, representing 2.96% of the study area. The classification performance of the best model was compared to the estimations of burned area done by the Chile’s National Forest Corporation (CONAF) (https://sit.conaf.cl/, accessed on 23 September 2021) and the FireCCI product [54].

4. Results and Discussion

4.1. Separability of Indices and Bands

Figure 4 shows that the highest values for the SI index occur mainly for the DELTA dataset. This demonstrates the advantage of using a multitemporal strategy to improve discrimination of the categories that have suffered spectral changes due to wildfires [9,55]. Furthermore, it is observed that for the DELTA dataset, the higher values are obtained by the spectral indices and not by individual bands. This shows the higher sensitivity of spectral indices that are designed to enhance the signal of burned areas [56]. As regards individual bands, those within the NIR zone (b6, b7 and b8A) show high separability values, which occur due to the great difference between the reflectance of vegetation affected or not affected by fire [4,57]. The NIR area is particularly sensitive to the internal cellular structure of leaves, and it reflects changes when an alteration takes place in this structure due to fire, senescence, a pathogen or deficiency of macronutrients [39]. On the contrary, bands of the visible zone (b2, b3, b4) show poor discrimination power, concurrent with what was obtained in [56]. Given the discrimination criteria considered, the features selected for training were the b6, b7 and b8A bands, and all of the spectral indices.

4.2. Performance Statistics

4.2.1. POST Dataset

Table 3. Best model performance for the POST dataset applied over the validation set.

Algorithm	DICE	Accuracy	Omission	Commission
RF	0.92	0.93	0.08	0.09
SVM	0.87	0.88	0.16	0.09
LR	0.88	0.89	0.14	0.11
MLP	0.87	0.87	0.10	0.14
ELM	0.92	0.92	0.02	0.12

where RF: random forest; SVM: support vector machine; LR: logistic regression; MLP: multilayer perceptron; and ELM: extreme learning machine.

presents the performance of the best model obtained from the POST dataset for the validation set. On the other hand,

Table 4. Best model performance for the POST dataset applied over the Test set.

Algorithm	DICE	Accuracy	Omission	Commission
RF	0.93	0.92	0.08	0.08
SVM	0.86	0.88	0.12	0.10
LR	0.88	0.88	0.12	0.10
MLP	0.80	0.85	0.01	0.12
ELM	0.89	0.89	0.01	0.11

shows the best model performance of POST data for the test set. The highest performance in both tables was obtained by RF, followed by ELM, SVM, LR and MLP. As was expected, the test set experienced a low decrease in DICE and accuracy performance indicators, but an improvement for omission and commission errors.

4.2.2. DELTA Dataset

Table 5. Best model performance over the DELTA dataset for the validation set.

Algorithm	DICE	Accuracy	Omission	Commission
RF	0.94	0.94	0.08	0.06
SVM	0.89	0.89	0.16	0.08
LR	0.84	0.84	0.11	0.16
MLP	0.89	0.89	0.04	0.12
ELM	0.93	0.92	0.12	0.02

presents the best model performance for the validation set and

Table 6. Best model performance over the DELTA dataset for the test set.

Algorithm	DICE	Accuracy	Omission	Commission
RF	0.92	0.93	0.07	0.08
SVM	0.87	0.89	0.16	0.08
LR	0.84	0.83	0.17	0.19
MLP	0.86	0.85	0.10	0.10
ELM	0.91	0.91	0.12	0.02

the best model performance for the test set. It is observed that the classifiers have better performance for the DELTA dataset than for the POST dataset, seemingly due to the fact that DELTA shows higher levels of SI in almost all the features analyzed (Figure 4). Moreover, the order of performance of the algorithms is maintained for the POST dataset, showing a similar decreasing tendency for the test dataset indicators.

In [7], similar results were obtained when RF was compared to SVM and MLP over MODIS data with 500 m resolution. In this study, the worst performance for most algorithms occurred for the POST dataset. This may have been caused by the high variability of the burned class [58]. On the contrary, this variability was considerably reduced for the DELTA dataset, which accounts for the better performance of the algorithms [52]. The problems mentioned in the POST dataset could be reduced, including a larger amount of training samples of different land covers and burned zones with different conditions [59]. Although the data used for the training sets were thoroughly revised, a high number of pixels with atypical values were found, which could have been wrongly classified. This highlights the importance of adopting a strategy of several steps where the results obtained are refined through the elimination or inclusion of training areas in the most conflicting zones [59].

Mostly, omission errors had a greater magnitude over the DELTA sets than over the POST sets. The opposite took place for commission errors. Omissions can be explained by the high variability of burned areas, which present zones of different levels of severity and cover, such as shrublands and agricultural burning. Some of them present both low contrast with their unburned environment and areas in which vegetation has begun recovering after the fire [9]. On the other hand, commission errors were caused by the topographic shading that was not completely removed by the correction conducted, the presence of elements of great height (trees that cast shadows on border areas) [59], areas with scarce vegetation and high presence of soil (agricultural and forest harvested), whose spectral changes are consistent with burned areas [60].

4.3. Time Statistics

Figure 5 shows the training time of the best model of the classifiers used for both POST and DELTA datasets. ELM (1.45 s) and LR (1.85 s) algorithms recorded the shortest training times, followed by SVM (6.2 s) and RF (35.1 s). In line with the hypothesis, ELM neural network training times are shorter than the rest of the classifiers. This is due to the fact that the training algorithm does not optimize all the network parameters and the calculation of the Moore–Penrose pseudoinverse is highly efficient [61].

4.4. Comparison of the ELM Classification with Burned Area Products

This section presents a comparison with methods of estimation of burned areas operatively used, namely, the method used by CONAF [62] and the FireCCI algorithm [54].

The best indicators in the validation area were reached by ELM, followed by CONAF and FIRECCI (

Table 7. ELM–CONAF–FIRECCI comparison.

-	Ground Truth	ELM	CONAF	FIRECCI
Burned area (ha)	37,472.28	39,702.96	50,216.84	43,868.44
Unburned area (ha)	209,928.72	207,698.04	197,184.16	203,532.56
DICE	1	0.857	0.799	0.773
Omission	0	0.080	0.064	0.160
Commission	0	0.132	0.301	0.282

). The higher spatial coincidence (DICE = 0.857) occurred between the classification carried out by ELM and the reference one, resulting in low omission (0.080) and commission (0.132) errors. The spatial coincidence of ELM classification with the reference data is observed in the detailed description that the ELM carries out over the burned-area borders and in the capacity to describe the internal islands that were not affected by fire (Figure 6D). This was explained by the proper training of the network, its generalization capacity, and the correspondence with the spatial resolution of the images used. On the other hand, estimates by CONAF showed lower spatial coincidence with respect to the classification of reference, having a DICE of 0.799, with omission and commission errors of 0.064 and 0.301, respectively. Despite showing a good spatial coincidence with its reference, and using Landsat 7 and 8 images [62], the estimate by CONAF showed a higher number of commission errors. They can be presumably caused by a general visual interpretation of the areas affected by fire, making it hard to account for the complexity of the geometry of the burned polygons and the unburned islands (Figure 6B). Moreover, FIRECCI estimations showed the worst results for the validation metrics. The low spatial resolution of the FIRECCI product (250 m) prevents it from detecting internal unburned islands, which are located to the interior of the big burned patches (Figure 6C). Furthermore, it prevents it from detecting low-intensity fires, mainly of agricultural origin [13].

5. Conclusions

This study has evaluated the potential of ELM neural networks as a new burned area classifier of multispectral Sentinel-2 satellite images. ELM showed promising results in the classification performance on the test dataset and shorter training times than the other algorithms. This characteristic is a great advantage for ELM in order to address burned area mapping with medium-high spatial resolution images at national and global scales. However, like the other machine-learning algorithms, its practical implementation requires preprocessing stages that include an atmospheric and topographic correction and samples collection to generate training datasets with enough spectral and spatial representativity of the burned areas.

In the current context of climate change, where an increase in the frequency and magnitude of wildfires is forecasted, there is a growing need for precise information about burned areas. This information is essential to fire risk systems and burned-area records used to design prevention and fire-combat strategies and provides valuable knowledge on the effect of fires on the landscape and the atmosphere.