Science of Remote Sensing 8 (2023) 100101

Contents lists available at ScienceDirect

Science of Remote Sensing

journal homepage: www.sciencedirect.com/journal/science-of-remote-sensing

Improving wildland fire spread prediction using deep U-Nets

Fadoua Khennou, Moulay A. Akhloufi *
Perception, Robotics and Intelligent Machines, Université de Moncton, Moncton, E1A 3E9, Canada

A R T I C L E I N F O A B S T R A C T

Keywords: Forest fires are able to cause significant damage to humans and the earth’s fauna and flora. If a fire is not
Forest fires detected and extinguished before it spreads, it can have disastrous results. In addition to satellite images, recent
Fire spread modelling studies have shown that exploring both weather and topography characteristics is crucial for effectively pre­
Deep learning
dicting the propagation of wildfires. In this paper, we present FU-NetCastV2, a deep learning convolutional
Convolutional neural networks
neural network for fire spread and burned area mapping. This algorithm predicts which areas around wildfires
are at high risk of future spread. With an accuracy of 94.6% and an AUC of 97.7%, the model surpassed the
literature by 3.7% and exhibited a 1.9% improvement over our previous model. The proposed approach was
implemented using consecutive forest wildfire perimeters, satellite images, Digital Elevation Model maps, aspect,
slope and weather data.

1. Introduction the slope, which has a direct effect on the sun exposure and the moisture
content of the vegetation NASA’s; Cal; Estes et al., (2017). Besides these,
Worldwide, we have lost almost 100 million hectares of forests in the wind can increase the speed of propagation by a factor of 100 and it is
last two decades thus accelerating the loss of biodiversity World; considered one of the most influential parameters of a fire’s propagation
Wildfire. Firefighting professionals have recently noticed that as fires get Rossa and Fernandes (2018); Eftekharian et al. (2019).
more numerous and intense they are becoming increasingly difficult to In order to tackle the significant losses of global forest resources,
extinguish. Between 2012 and 2021, there were 61,289 wildfires on which in turn have a large impact on the global ecological balance,
average a year, affecting an average of 7.4 million acres according to prevention is the ideal solution. In this context, artificial intelligence
statistics by the Congressional Research Service Boer, de Dios and (AI) can be used to identify the conditions and the circumstances of a fire
Bradstock (2020); Hoover and Hanson. outbreak along with its evolution over time Imada (2014). AI can also be
There are three elements that must be present for a fire to start; heat, used to coordinate interventions in the field by indicating certain areas
fuel and oxygen Güngöroğlu (2018). Together, they form what is called to prioritize as well as when it should be prioritized Calp and Kose
the ”fire triangle”. The first element is the heat, which is the spark or the (2020); Al-Kahlout et al. (2020). Just as humans learn from experiences,
source that brings fuel to its ignition temperature. It’s source can be AI models learn from past events. Common algorithms often taken as
human, such as a lit match, or natural, such as lightning strikes. The input a history of data describing the characteristics of past fires as well
second element is the fuel, which in our case is vegetation, for example as the weather and vegetation conditions associated with them. The
trees, dead leaves and twigs on the ground, or other flammable material. algorithm will then look for correlations between these environmental
Lastly we have oxygen, which is one of the main components of the factors and the wildfire perimeters of a specific region Al-Kahlout et al.
atmosphere and which is required for combustion. In order to control or (2020).
extinguish a fire, at least one element of the fire triangle must be elim­ The intersection of deep learning and wildfire science covers a wide
inated Alkhatib and Szmyt (2017). spectrum of subjects Jain et al. (2020); Song and Wang (2020). Although
Vegetation is crucial aspect of a fire’s ability to spread. Depending on the use of neural networks, particularly convolutional neural networks
the species, its different chemical and physical properties directly in­ (CNNs), has been trending in various wildfire detection and prediction
fluence the spread. In addition to that, topography has a significant in­ approaches, there have been relatively few studies that include tem­
fluence on the propagation of fire thanks to multiple factors. Such perature, topography, and satellite data to map predicted burned areas
factors include altitude, which has an impact on the oxygen level, and Liang et al. (2019); Radke et al. (2019); Elia, D’Este, Ascoli, Giannico,

* Corresponding author.
E-mail addresses: [email protected] (F. Khennou), [email protected] (M.A. Akhloufi).

https://doi.org/10.1016/j.srs.2023.100101
Received 27 February 2023; Received in revised form 23 May 2023; Accepted 11 September 2023
Available online 15 September 2023
2666-0172/© 2023 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

Spano, Ganga, Colangelo, Lafortezza and Sanesi (2020). Fire or smoke LiDAR point clouds. Data from three case studies with diverse species
detection and recognition systems are being given increased attention compositions, densities, homogeneities, and ages were collected using a
Abid. While detection systems are undoubtedly very useful and neces­ UAV platform. In comparison to the naturally regenerated samples,
sary as they can detect fires immediately after or before they have where the results varied between 58% and 49%, trees detections accu­
ignited, wildfire mapping models can further assist firefighters in un­ racy was quite high for the plantation sample. They concluded that the
derstanding the behavior of forest fires using different characteristics findings approve the impact of the homogenous species composition,
and features. regular individual distribution pattern, and conical form of the crowns.
The fire spread model described in this article aims at addressing the The non-conical shaped crowns were used to explain lower accuracy
problem of forest fire propagation after a 24 h period using consecutive findings in other datasets.
fire perimeters, satellite, topographic, and climatic features from the In another study Wulder et al. (2009), multi-temporal Landsat was
Rocky Mountains. used alongside LiDAR to investigate on the consequences of post-fire
This work is organized as follows. Section 2 describes current effects. Image segments were generated using the post-fire Landsat
research findings that significantly improved the literature on wildfire data in order to provide a geographical framework for contrasting the
detection and propagation models. The suggested architecture, mate­ traits and temporal dynamics of the forest structure. The study’s findings
rials, and methods are all described in Section 3. The experimental re­ demonstrated how LIDAR data can be used to characterise post-fire
sults are shown in Section 4. Finally, the conclusion is presented in conditions and burn-induced structural change. These data can be
Section 5. used in conjunction with other factors, such as vegetation type, mois­
ture, topography and long-term weather patterns.
2. Related works In the context of supervised and unsupervised machine learning, the
researchers in Farhani et al. (2021), implemented different approaches
2.1. Wildfire detection for level-0 lidar data classification and clustering. They examined the
ability of machine learning to detect anomalies (traces of wildfire smoke
Nowadays, forest fires can be detected using a variety of automated in lower stratosphere) and proposed a classification-based model for
monitoring tools. Satellite systems and images, wireless sensor networks lidar measurement profiles. Gradient boosting and random forests ach­
and infrared (IR) detectors are among the most commonly used tech­ ieved the best performance with a higher accuracy of 98%. As for pro­
nologies and systems for this Hristov et al. (2018); Dampage et al.. Fig. 1 files’ clustering with unsupervised learning, using the t-distributed
presents the modelling process of a wildfire detection system using deep stochastic neighbour embedding (t-sne), the model effectively clusters
learning. profiles on nights with both constant and changing lidar profiles
A recent research review summarized the use of unmanned aerial dependent on climatic variables and system alignment.
vehicles (UAVs) in the context of fire detection and monitoring Akhloufi For deep networks, CNNs have recently been widely applied to the
et al. (2021). The researchers presented an overview of sensing in­ problem of fire detection as they are capable of extracting features and
struments, fire perception algorithms and different coordination stra­ patterns from spatial images and so they are being extensively used in
tegies. They concluded that these technologies and specifically object recognition applications Simard et al. (2003); Park et al. (2020);
cooperative autonomous systems can play an important role in the fight Sayad et al. (2019); Wang et al. (2019); Muhammad et al. (2018); Li
against wildland fires in large areas. et al. (2019). Many recent applications are trained using forest fire or
On the other hand, LiDAR (Light Detection and Ranging) technology smoke images. Sayad et al. prepared a fire dataset using MODIS (Mod­
has made significant advancements in a variety of fields, most notably in erate Resolution Imaging Spectroradiometer), a sensor embedded in
forest fires prevention and detection. They are commonly used in at­ both Terra and Aqua satellites Sayad et al. (2019). The goal of this study
mospheric research and offer excellent temporal and geographical res­ was to study wildfire occurrences in a specific region of Canada’s forests
olution profiling Hopkinson et al. (2004). between 2013 and 2014. Using Spark, a Big Data framework, they
The researchers Fernández-Álvarez et al. (2019), presented a meth­ implemented both Support vector machines (SVMs) Boser et al. (1992)
odology that enables automated evaluation of vegetation’s adherence to and Neural Networks which achieved an accuracy of 97.48% and
geometric limits that were set in place for wildfire prevention. To in­ 98.32% respectively. They also optimized their models using cross
crease the wildland urban interface and infrastructures’ protective ca­ validation and regularization techniques. Another research experiment
pacities, they characterised the forest vegetation using high-resolution was conducted using a CNN based on Inception-v3 Szegedy et al. (2015);
Alves et al. (2019). This deeper network architecture, with an input size
of 299 × 299, needs fewer parameters and less time to train. Its goal was
to classify fire and non-fire images. The weighted average for their fire
detection accuracy was 98% using 481 images for training and 53 im­
ages for testing.
Jiao et al. Jiao et al. (2019) proposed an approach that can be applied
on forest fire prevention with UAVs. Their method uses aerial images
and the YOLOv3 algorithm. One of limit of their model is that it was only
tested on a small dataset. They achieved a detection rate of 83%.
In order to address the limitation of smaller datasets and following a
prior analysis of fire detection solutions, the researchers in Alves et al.
(2019) prepared a dataset containing 882 images which covers a range
of real world scenarios with images from a multitude of landscape re­
gions in different seasons. They concluded that despite attaining a high
performance detection rate, prior investigations had a major drawback
because of the image quality of the trained fires. Using Inception-V3, the
models using daytime and nighttime scenarios were both evaluated and
found to be 94.1% and 94.8% accurate respectively. Their results
demonstrated that the most significant feature for the occurrence of false
positives by the classification model was the presence of fog during the
Fig. 1. Wildfire detection system using deep learning. day and artificial lights at night. This showcased that more studies need

2
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

to test the performance of the detection of small fire spots using a vol­ and heterogeneous landscapes achieving an F-measure of 0.93. Their
uminous training set. approach was able to predict future maps up to 24 h in advance.
As for early forest smoke recognition, the researchers in Cao et al. In Nambia, based on a 16-year record (2000–2016) of the MODIS
(2019) developed an attention enhanced bidirectional long short-term Burned Area, Mayr et al. examined the biophysical and human-related
memory (LSTM) network (ABi-LSTM) Hochreiter and Schmidhuber controls of fire activity Mayr et al. (2018). The burned area along with
(1997) with which they achieved an accuracy of 97.8%. They extracted the number of fire occurrences are two derived metrics of fire activity
spatial data from smoke location candidates using the Inception V3 that are explored separately at a 0.1 resolution using five popular sta­
network and then applied the BiLSTM model which extracts temporal tistical and machine-learning techniques. The findings, based on
features in forward and backward order by feeding spatial characteris­ Random Forests modelling results, show the relevance of mean annual
tics from these patches. Finally, with a soft attention mechanism that precipitation and human influences as factors impacting fuel availabil­
automatically measures the relevance of distinct frames, an attention ity. Finally, they attributed the presence of moderate green-up rates of
network is used to optimize the classification process. To increase the vegetation in the most burned areas to occurrences of open landscapes.
performance of the previously proposed models, this study suggests that Mithal et al. presented a burned area forecasting system that lever­
heterogeneous environments must be used. As a result, it is crucial to ages information from MODIS active fire hotspots and multispectral
look into extended CNN architectures that could help improve existing surface reflectance data for global-scale burned area detection Mithal
results. et al. (2018). The predictive model is trained on land cover and
geographical locations with characteristics similar to the tropical forests
2.2. Wildfire burned area mapping in South America and South-east Asia. They used neural networks to
model the probability that a given pixel belongs to the burned or
Wildfire spread rate is determined by an interaction between fuels, non-burned class. They compared their model to reference validation
weather and topography. A brush fire is affected by the interactions of maps created from Landsat images and observed that it identifies about
these three factors Güngöroğlu (2018). The size of the fuel affects how 60% of burned areas which were reported by MCD64 Giglio et al.
quickly it burns and fuel arrangement, the horizontal and vertical dis­ (2015).
tribution of all flammable materials, also affects the burning rate. In another researchBjanes et al. (2021), Bjanes et al. considered
As for weather, there are different meteorological factors which fifteen fire influencing factors categorized into climatic, topographic,
directly influence the behavior of a fire, such as relative humidity, vegetation-related and anthropogenic. From the probability maps pre­
precipitation, temperature and wind direction Koutsias et al.; Zheng dicted by their ensemble learning and CNN models, two observations
et al.; Martell and Sun. Relative humidity, the amount of moisture in the were drawn about the predictor variables. The first considers that in­
air, affects the moisture content of the fuel. Precipitation, including rain, tensity of the probabilities between January and July can be attributed
hail and snow, also affects the fuel’s moisture content. Lastly, wind is to the dynamic climatic variables, suggesting that the meteorological
one of the most important factors determining the behavior of fire, and the vegetation’s condition, have an impact in how the models
affecting both the direction and the rate of a brush fire’s spread Ma et al. discriminate fire class from no-fire class. The second, highlights that
(2020). By directing moist air either toward or away from the fuel, wind static variables, i.e., topographic and anthropogenic could be considered
strongly impacts how quickly it dries out and can even fan embers that as predisposing factors for ignition. A study of the probability that fires
start new fires elsewhere Holsinger et al. (2016). would occur on a scale from January to July served as the basis for their
Topography is the set of physical characteristics of the ground sur­ prediction scale.
face. These characteristics are evaluated based on whether they’ll stop, Researchers in Shmuel and Heifetz (2022), compared different ma­
counter or promote the fire’s spread Ahmad et al. (2018). The slope chine learning algorithms including random forest, XGBoost, multilayer
represents the change in elevation over a specific distance. When a fire perceptron, logistic and linear regressions. Their study was based on a
spreads up a slope, the fuel is more easily heated, which means that fires one year scale wildfire observations from around the globe, achieving
spread faster going uphill. The exposure is the direction in which the 90% accuracy with XGBoost. Once again, their study highlights the
slope is oriented. South-facing slopes are the most exposed to the sun, importance of regional wildfire history alongside weather data for the
making the fuel hotter and drier, and therefore easier to burn Guo et al. prediction of fire occurrences.
(2016). In the same context, other researchers have been much more inter­
In order to minimize wildfires damage using fire detection systems, it ested in the prediction of fire occurrences rather than the prediction of
is important to instantly monitor a fire’s spread to predict its progression wildfire spread after its ignition Zhang et al. (2021); Bergado et al.
over a time interval. This can ultimately give information to the fire (2021); Naderpour et al. (2021). We believe this is due to the complexity
fighters so that they can act quickly and efficiently by deploying the of gathering relevant data for the prediction. Notably, since fire spread
right measures in a timely manner. While rapid detection and prediction modelling requires different features for the same area of interest,
systems have become crucial, traditional monitoring tools, e.g. using including both initial ignitions and progressive mapping.
human observation towers, can no longer meet our needs as their The convergence of Artificial Intelligence decision-making tools,
effectiveness is limited by temporal coverage and human error Jain et al. deep learning models and challenging wildfire datasets provides a rich
(2020). Therefore in these circumstances the use of artificial intelligence ground for research. It is clear that forest fire modelling is a complex
techniques has become imperative in order to process and analyze a process as its occurrence and behavior are the result of a number of
variety of forest fire data and propose appropriate tools for fire spread interconnected causes, notably the source of ignition, the nature of the
prediction and specifically for modelling fire growth Tonini et al., fuel, the weather and the topography Zhao et al.; Kinaneva et al. Fig. 2
(2020); de Bem et al. (2019). summarizes the modelling process of a wildfire spread prediction system
Researchers are developing new models that are able to study and using deep learning.
analyze the behavior of forest fires to better predict future burned areas Following a thorough analysis of the present literature, we observe
and to classify them according to severity Radke et al. (2019); Khennou that there is still a gap in this particular area of spread modelling. This is
et al. (2021); Liang et al. (2019); Subramanian and Crowley (2017); unlike detection systems which have been widely implemented using
Hodges and Lattimer (2019). For burned area modelling, Jonathan et al. different neural networks. To this end, we provide evidence for an
Hodges and Lattimer (2019) proposed a machine learning approach to affirmative solution to this problem by introducing a new approach
estimate the spatial evolution of a wildland fire front using a deep based on U-Net to automatically learn and understand wildfire spread
convolutional inverse graphics network (DCIGN). The particularity of dynamics.
their approach is that they tested their model using both homogeneous The current study’s key contribution is the integration of weather,

3
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

Fig. 2. Wildfire spread prediction system using deep learning.

topography, satellite and previous fire perimeters of the Rocky Moun­

tains to forecast the spread after a 24h period. U-Net, a fully convolu­
tional network Ronneberger et al. (2015), is optimized and implemented
for the purpose of this study. This work is an extended version and an
optimization of a FU-NetCast base model Khennou et al. (2021). This
contribution consists of the acquisition of new training data and an
optimization of the obtained results using U-Net, which is considered Fig. 3. Example of a fire perimeter progression over 24 h (fire number 33 from
one of the best neural networks for precise image segmentation Wang GeoMAC2019 pairs). The size of the area shown is approximately 28 × 20 km.
et al.; Ghali et al.; Frizzi et.al.

3. Methods

In this section, we describe our materials and methods. This includes

the datasets we used, the proposed model architecture along with the
evaluation metrics.

3.1. Fire perimeter

We used GeoMAC data from 2013 to 2019 to get fire perimeters from
the NIFC (Historic wildfire perimeters). We downloaded each year’s
shape file and spreadsheet and implemented a script to retrieve the pairs
of successive perimeters separated by a 24-h delay. Fig. 3 presents an
example of a fire’s perimeter progression over 24 h.
The goal behind the 24-h scale is to implement a deep learning model
that is able to forecast the fire spread using an initial burned area and
analyze how satellite images and continuous time series weather data Fig. 4. Digital elevation model.
impact the final perimeter of the wildfire.

P − min(image)
P′ = (1)
3.2. Digital elevation model max(all images) − min(all images)
In fact, the learning process can be disrupted or slowed down when
With a resolution of 1 arc-second, the Digital Elevation Model (DEM)
high integer input values are used in neural networks, which analyze
is obtained from the USGS 3D Elevation Program (3DEP). The whole
inputs using modest weight values. In this vein, it is crucial to normalize
directory of the National Map, which covers North America, is down­
the pixel values such that each one has a value between 0 and 1. The
loaded from Data.Gov. Fig. 4 presents an example of a DEM for a specific
potential minimum and maximum values supported by the data are
region.
represented using the min() and max() values.
From the DEM, the slope (landscape slope) and the direction (car­
dinal angle) of the slope (landscape aspect) are retrieved. The challenge
when working with the slope parameter is that it affects both radiation 3.3. Satellite images
and convectional heat transfer (due to the interaction between topog­
raphy and the flame front). In terms of the landscape’s orientation or Landsat-8 data was downloaded in the GeoTIFF format from Google
aspect, several factors can affect how fires burn. Notably, they have an Cloud Storage using the gsutil tool with a resolution of 30 m Survey
impact on the types of plants that grow there as well as how much (Survey).
moisture they hold. In order to consider the effect of vegetation a fuel source, we use
For the DEM normalization, we apply the following formula to each satellite images to provide us with information about the vegetation
pixel P: state as this information is otherwise impossible to gather on a large

4
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

scale. Several frequency bands are accessible for each case, as shown in 3.4. Meteorological data
Fig. 5. The red, green, blue, and IR bands (2-3-4-5) were used by Radke
et al. Radke et al. (2019) because they suggest that these bands correlate Environmental circumstances change over time, even within a day,
with vegetation. Bands 6 and 7 may, nevertheless, be useful as well. It is and these changes have significant impact on the water content of
crucial to take into consideration that satellite images are the only input vegetation. On a hill, the spread of fire can be up to 20 times faster than
variables that can account for physical impediments like roads and on flat ground. The effect of the wind, on the other hand, can boost the
rivers. It is logical to assume that the presented bands (especially those propagation speed by a factor of up to 100 Rossa and Fernandes (2018);
in the visible spectrum) will recognise them if they are wide enough. Eftekharian et al. (2019).
An estimate of the Area of Interest (AOI) is defined for each pair of We downloaded data from the Visual Crossing Weather API Crossing.
perimeters with a rectangle with a 2 km perimeter. Because this con­ Hourly meteorological data between the initial and final perimeter as
dition leads some pairs to be rejected, we use scenes that completely well as daily data for the 25 days preceding the fire are obtained for a
cover the AOI in a single image. We also choose the most recent scene point in the center of the AOI to represent the entire region as described
prior to a fire that had less than 10% cloud cover across the entire scene in Table 1. After all the data is prepared, it is scaled from 0 to 1. Linear
(max cloud). scaling is calculated separately for each feature.
We implemented a python script to re-project the images (DEM,
perimeters) in the satellite image’s Coordinate Reference System (CRS)

Fig. 5. Landsat 8 operational land imager (OLI) and thermal infrared sensor (TIRS).

5
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

Table 1 data point of the corresponding fire perimeter.

Feature characteristics. The burned area prediction framework makes use of the information
Features Source Dimension from fire perimeters, weather, topography and elevation model data
previously described (see Fig. 7). The proposed methodology, schema­
DEM USGS 3D Elevation Program 30 × 30 m
Slope USGS 3D Elevation Program 30 × 30 m tized in Fig. 6, is described as follows.
Aspect USGS 3D Elevation Program 30 × 30 m
Satellite bands 2-7 Google Cloud (Landsat Data) 30 × 30 m 1. Collecting and gathering historic GeoMac fire perimeters, elevation
Early precipitation Visual crossing Numeric and sattelite maps.
Relative humidity Visual crossing Numeric
Wind NS Visual crossing Numeric
2. Retrieving 24 h weather data related to the collected fire perimeters.
Wind EW Visual crossing Numeric 3. Preprocessing and preparing data.
Temperature Visual crossing Numeric 4. Implementing and optimizing U-Nets.
Sea Level Pressure Visual crossing Numeric 5. Generating the final burned area map.
Prec 1 week before Visual crossing Numeric
6. Evaluating the results based on the Area Under the Curve (AUC),
Prec 2 weeks before Visual crossing Numeric
Prec 25 days before Visual crossing Numeric accuracy and F-measure metrics.
RH 1 week before Visual crossing Numeric
RH 2 weeks before Visual crossing Numeric FU-NetCastV2 is a wildfire burned area prediction model that uses a
RH 25 days before Visual crossing Numeric modified version of the U-Net architecture. It is specifically designed to
analyze 24-h time series data and provide accurate predictions of the
and to normalize the scale of each pair of perimeters (1 pixel = 30 m). extent of burned areas caused by wildfires. FU-NetCastV2 builds upon
Furthermore, the DEM, satellite bands 2–7, and initial and final perim­ the architecture of U-Net and tailors it to the specific requirements of
eters are cropped for each pair of perimeters to cover 2 km around the wildfire burned area prediction. The model takes into account the
burned zone. temporal nature of the data, allowing it to capture the dynamics and
patterns of fire propagation over a 24-h period. We can summarize the
architecture in Fig. 6 as follows.
3.5. Approach
⋅ Input Data: FU-NetCastV2 takes as input a sequence of satellite
The U-Net architecture is based on the concept of the fully con­ covering a 24-h period. These images typically capture different
volutional network Ronneberger et al. (2015). The fundamental idea is spectral bands, such as near-infrared, red, and green, which provide
to add layers to a traditional contracting network and replacing pooling valuable information about vegetation and fire intensity.
operations with upsampling operators. As a result, these layers improve ⋅ Encoding Path: The model begins with an encoding path that consists
the output resolution. Furthermore, depending on this knowledge, a of a series of convolutional layers. These layers perform feature
subsequent convolutional layer can learn to calculate an accurate extraction on the input images, gradually reducing their spatial di­
output. In a classical convolutional neural network, a successive mensions while increasing the number of channels to capture higher-
reduction in image dimensions is necessary for the extraction of relevant level representations.
information and for generalization. However, the location information is ⋅ Skip Connections: To preserve spatial information, skip connections
lost. This is where semantic segmentation algorithms differ. are introduced between the encoder and decoder paths. These con­
U-Net was developed for biomedical image segmentation in 2016 nections allow the model to combine low-level spatial details from
Ronneberger et al. (2015). This architecture was later demonstrated to the encoder path with high-level contextual information from the
be useful for satellite images applications Guo et al.; Sofla et al.. It has decoder path.
the benefit of being theoretically simple, as well as having the ability to ⋅ Decoding Path: The decoding path starts with upsampling operations
process images of various sizes and with as many bands as necessary. to gradually restore the spatial dimensions of the encoded features.
There are two phases to the architecture. The first is the contraction These operations are typically performed using transposed convo­
phase, often known as the encoder, which records the image’s context lutions or interpolation techniques.
and data. It uses combination of convolutional and max pooling layers. ⋅ Concatenation: At each upsampling step, the decoder path concate­
The second phase is the expansion phase, which employs transposed nates the upsampled features with the corresponding features from
convolutions (or up-convolutions) to provide exact image pattern the skip connections. This enables the model to incorporate both
localization. local and global information, facilitating the accurate prediction of
We combine each result of the transposed convolutions, operations burned areas.
which increase the size of the image, with the same-sized images from ⋅ Final Prediction: The output of the decoding path is passed through a
the encoding phase during the expansion. This allows for the spatial convolutional layer with a sigmoid activation function. This gener­
information to be reintroduced step by step. The transposed convolution ates a pixel-wise prediction map, where each pixel represents the
(up-convolution) involves multiplying the value of each source pixel probability of being a burned area.
with a mask and replicating the result in a bigger output matrix with a
space between each copy, allowing for an overlap if the gap is smaller Each block is created based on a convolutional layer with a 3*3 filter
than the mask size. For each pixel, U-Net employs a loss function. This and a stride of 1. It is then followed by a batch normalization, which
makes it easy to distinguish pixels with burned areas from pixels without helps in normalizing the layer’s output and stabilising the model’s
fire in the segmentation map. learning process. Starting with 64 channels and as the network gets
One of the unique features of the Bottleneck layer in U-Net is that it deeper, the number doubles after each max pooling operation and rea­
helps in minimizing the mutual information for the displayed layers, ches 1024 in the bottleneck. Here, the result of the transposed convo­
identifying the mapping that contains only the relevant information, and lution is combined with the output of the matching decoder portion and
so identifying the true existing relationship in the data. Additionally, in a numerical layer, where we input weather data.
comparison to CNNs, the implementation of Skip-connections in U-Net, Initially, we used RELU for training, but after a few iterations, the
helps in concatenating the feature maps from the contraction (encoder) model started to predict the same outcome for every batch input and the
to the expansion path (decoder), providing the network with additional RELU was unlikely to recover once it reached this stage. Hence, we
localization data about the image. Hence, we exploited this localization employed the Leaky ReLU, which is the most effective function for
to input numerical features that were combined respectively to each solving a dying RELU.

6
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

Fig. 6. Architecture of the FU-NetCastV2 model.

We implemented a forward() function inside the Unet class, which

takes as input eleven feature maps (aspect, initial burn, dem, satB2,
satB3, satB4, satB5, satB6, satB7, slope, final burn) and appended a
numerical data channel, which includes an array of twelve weather data
(temperature, precipitation, relative humidity, wind EW, wind NS,
pressure, prec 1week before, prec 2weeks before, prec 25days before,
RH 1week before, RH 2weeks before, RH 25days before). Here, we
structured and stored each of the first six features of weather data into a
two-dimensional array, as these latter include 24 h data points, then we
concatenated the results with the remaining features.
For each pixel in the represented maps, Unet applies a loss function.
This facilitates the quick identification of particular burn locations on
the segmentation map. Each pixel is given the Softmax then a loss
function.This converts the segmentation into a classification problem
where we need to classify each pixel as a burn or not, hence the gen­
eration of the final map.
We used the Binary Cross Entropy with Logits Loss (BCE With Logits
Loss) as the network training loss function. The Sigmoid layer and Bi­
nary Cross Entropy Loss (BCE) are combined in the BCE With Logits Loss
function. As a result, the output is automatically processed via the Sig­
moid layer before being subjected to binary cross entropy computation
Fig. 7. Input maps of the FU-NetcastV2 model.
on both the output and the ground truth of the final burn map.

7
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

3.6. Evaluation metrics

PrecisionxRecall
F1score = 2X (6)
Precision + Recall
In this work, the following metrics will be used to evaluate and
analyze the Fu-NetCastV2 model.
3.6.5. Accuracy
The most common metric for classifier performance is accuracy. This
3.6.1. Sensitivity
statistic is simple to comprehend, yet it completely ignores several ele­
The proportion of positive burned points that are effectively identi­
ments that a model introduces into the system such as false positives and
fied is known as sensitivity. In other words, sensitivity refers to how well
false negatives. In our case study, we used the pixel accuracy to report
a test performs when applied to positive cases. A false negative rate can
the percentage of the correctly classified pixels in the burned fire images
be used to express this. The sum of sensitivity and false negative rate
Minaee et al.. This metric is generally used to evaluate a binary mask
would be 1. A higher sensitivity represents a greater true positive rate
where the true positive represents the pixel that was correctly predicted
and a lower false negative rate. It is calculated as shown in eq (2).
to belong to the given class and where a true negative represents a pixel
True Positive that was correctly identified as not belonging to the given class (in our
Sensitivity = (2)
True Positive + False Negative case whether is it burned or not).
TP + TN
3.6.2. Specificity Accuracy = (7)
TP + TN + FP + FN
The fraction of non-burned spots effectively recognised by the test is
known as specificity. In other words, the specificity of a test is a mea­ where:
surement of how well it performs when applied to negative cases. While TP is True Positive.
the sensitivity measure is used to determine the proportion of positive TN is True Negative.
cases that were successfully predicted, the Specificity measure is used to FP is False Positive.
determine the proportion of negative cases that were correctly pre­ FN is False Negative.
dicted. The metric is calculated as shown in eq (3). This statistic should only be considered if the classes are not signif­
True Negative icantly unbalanced. We used the AUC to further test our model since it is
Specificity = (3) one of the most common statistical approaches for estimating model
True Negative + False Positive
fitting.
3.6.3. Receiver operating characteristic curve (ROC)
The ROC curve corresponds to the graphic representation of the 4. Results
specificity and the sensitivity using different threshold values. It is
plotted as the true positive rate (TPR) against the false positive rate In this section, we present our model’s performance metrics and
(FPR), where TPR is on the y-axis and FPR is on the x-axis. The pa­ provide a comparative study in regards to recent research studies.
rameters of the ROC curve are presented as follows in eq (4) and eq (5): The receiver operating characteristic (ROC) was used to evaluate the
True Positive Rate (TPR) is a synonym for recall. It is a ratio of the models’ overall performance in addition to measuring their prediction
true positives to all positives. accuracy. The training dataset’s ROC represents the model’s success rate
/ and how closely the modelling results match the training dataset. Due to
TPR Recall =
True Positive
(4) memory constraints, a bigger batch size was not used for training the
True Positive + False Negative model. We trained for 100 epochs with a batch size of 1. We also used a
50% dropout on the input layer and shuffled the order of the samples
The false positive rate (FPR), is a measure of how often the positive
during training to avoid over-fitting. Furthermore, the learning rate was
class is incorrectly classified.
altered between 0.01 and 0.00001. Our model began to diverge at a
False Positive lower learning rate and training at 0.000001 was very slow. 0.00001
FPR = 1 − Specificity = (5)
True Negative + False Positive was the best-performing learning rate.
Based on our experiments, we can conclude that the model performs
The area under the ROC curve (AUC) is one of the most widely used
well and is able to accurately forecast the spread of wildfires based on
overall test performance measures Carter et al.. It varies from 0.5 for
time series weather data, satellite images, aspect, slope and elevation
non-informative tests to 1 for perfect performance. It is equal to the
maps. The model achieved an accuracy of 94.6%, an AUC of 97.7% and a
likelihood that a classifier would score a randomly chosen positive
F1-score of 95.9%. Table 2 shows the performance of our model on both
example higher than a randomly chosen negative example. In our case
training and validation sets.
study, it will measure how well the diagnostic test can distinguish be­
In Fig. 8 and Fig. 9, we present the predicted burned area for two
tween burned and unburned locations on the provided map.
sample maps in the Rocky Mountains. In the given maps, we trained our
model on the initial burn area, DEM, aspect, landsat bands (2–7) and
3.6.4. F1-score
slope. We then optimized the U-Net to train with combined weather data
As described in eq (6), the F1 score conveys the balance between the
such as precipitation, pressure, wind, relative humidity and sea level
precision and the recall and is equal to the harmonic mean of both
pressure with the second upsampling phase. These U-Net layers are
measures.
designed to boost the output’s resolution. The model’s high-resolution
Here, the precision is a metric that quantifies the proportion of
characteristics are mixed with the output for localization. The goal of
correctly predicted positive instances (true positives) out of all instances
the sequential convolution layer is to use this weather data to generate a
predicted as positive (true positives + false positives). It measures the
predicted burn map that is more exact. In order to improve the perfor­
model’s ability to avoid false positives.
mance of our model, we aggregated the weather data, presented in
As for the recall, it is also known as sensitivity or true positive rate,
Table 1, with 24 entries for each feature. These entries provide useful
measures the proportion of correctly predicted positive instances (true
information for climatic conditions throughout the day.
positives) out of all actual positive instances (true positives + false
In this context, the primary contribution of FU-NetcastV2 is the
negatives). It assesses the model’s ability to identify positive instances
concatenation of the higher resolution feature maps from the encoder
correctly.
network with the upsampled weather features when upsampling in
order to improve learning representations with subsequent

8
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

Table 2
Comparing the performance metrics on training and validation sets.
Data Accuracy Precision Recall F1-score AUC Sensitivity Specificity

Train 0.968 0.969 0.980 0.974 0.983 0.980 0.999

Test 0.946 0.947 0.970 0.959 0.977 0.970 0.984

convolutions. It it useful to have significant information in this step to

help with the localization since upsampling is a sparse operation. In
order to complement the fire perimeter in our input data, each burn day
from the training and testing fires is exposed to 24 ”future” hours of
atmospheric data. One particularity in our dataset is that it is quite
heterogeneous as it contains fire perimeters with both large and small
fire maps over 24 h intervals. In addition, we upsample features from the
bottleneck layer to ensure that the weather features are not down­
sampled and that each numerical value has its own significance, which
could be otherwise lost if included at the beginning of the U-Net model.
In our first training phase, we processed the model based on 100
epochs. After further analysis and optimizations, we noticed that the
model tends to overfit when it is training over 10 epochs. As presented in
Fig. 10, the model performs well as training and validation loss reach
their minimum values. The curves form a plateau as the number of
epochs increases. We considered and finetuned the architecture of Unet
as it is one of the most valuable architectures for pixel-based classifi­
cation. The down and upscaling architecture makes it a unique algo­
rithm that offers the potential for performing classification and
segmentation in a single step. Using this method, we assign a burn or no
burn classification to every pixel of the input maps to indicate the
presence of a fire. The reconstruction of images in the upsampling phase
of Unet (after a vector conversion) helps in restoring the image to its
original dimensions thus generate the final burn map as a predicted
output for our model.
Considering the attained metrics, we can conclude that the model is
able to generalize well even with unseen data. These results show that it
is possible to reduce the number of false alarms (to maximize speci­
Fig. 8. Initial and After 24 h burned maps for an area of interest n◦ 1. ficity). As a further evaluation, Fig. 11 displays the ROC curve of the test
dataset using the FU-NetcastV2 model.
Ground truth Vs. predicted burn (Area of interest 1) Ground truth Vs.
predicted burn (Area of interest 2).
Ground truth Vs. predicted burn (Area of interest 3) Ground truth Vs.
predicted burn (Area of interest 4).
In Fig. 12, we can see the shape of the burned map predicted by the
FU-Netcast model on the right compared to the ground truth on the left.
We notice that the represented results are highly consistent. In Table 3,
we present a comparative study of our model’s performance with

Fig. 9. Initial and After 24 h burned maps for an area of interest n◦ 2.

Fig. 10. Train and Validation Learning using FU-NetcastV2.

9
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

Fig. 11. ROC curve using U-Net.

Fig. 12. Sample maps representing the ground truth Vs. the 24h predicted burn using FU-NetcastV2.

existing studies which also used deep learning techniques in both outperformed the literature by a margin of 3.7% and exhibited a notable
wildfire detection and spread mapping models. 1.9% improvement compared to our previous version Khennou et al.
The researchers Liang et al. Liang et al. (2019) implemented a (2021). The GeoMAC data from 2013 to 2019 includes both initial and
backpropagation neural network (BPNN), a recurrent neural network after 24 h maps, representing the progression of wildfires. In addition to
(RNN) and long short-term memory (LSTM) network for wildfire spread that, fire progression distribution after this time slot can change dras­
prediction models. Their approach was a multi classification tool to tically, yet, we initially considered heterogeneities in the representation
predict the scale of the fire by classifying it within specific scale levels of our dataset. We included fire perimeters where there has been both
(from 1 to 5). They based their study on numerical inputs data, such as important progressions and small-scale ones providing an area of in­
latitude, longitude, date of ignition, date of extinction and meteoro­ terest, as the case of the represented maps in figure Figs. 8 and 9.
logical data and achieved the best performance using the LSTM network. This research also provides a significant performance improvement
In the same context of fire spread modelling, Radke et al.Radke et al. compared to the FU-Netcast model, which was implemented using 120
(2019) obtained an average accuracy of 87.7% using 9 fires from the images Khennou et al. (2021).
GeoMAC dataset using data augmentation to expand their training data Our technique is likewise comparable to some of the spread predic­
by a factor of 10. In contrast, FU-NetcastV2 achieved an accuracy of tion models like Radke et al.Radke et al. (2019) who also used the 24h
94.6%, an AUC of 97.7% and an F1-score of 95.9% using a dataset of 400 interval for the scale mapping. While the majority focused solely on the
images with each feature containing 400 maps and weather time series burned region Liang et al. (2019); Hodges and Lattimer (2019); Sub­
data using a 80% and 20% ratio for training and testing respectively. It ramanian and Crowley (2017), our approach responds to the more

10
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

Table 3
Performance Metric Comparison across Recent Studies in Wildfire Prediction and detection.
Model Approach Accuracy Precision Recall F1-score AUC

Spread CNN Hodges and Lattimer (2019) – 0.890 0.800 0.820 –

U-Net Khennou et al. (2021) 0.927 0.801 0.790 0.795 0.800
CNN Radke et al. (2019) 0.877 – – – –
BPNN Liang et al. (2019) 0.704 – – – –
RNN Liang et al. (2019) 0.835 – – – –
LSTM Liang et al. (2019) 0.909 – – – 0.942
RL Ganapathi Subramanian and Crowley (2018) 0.901 – – – 0.929
FU-NetcastV2 0.946 0.947 0.970 0.959 0.977

Detection DenseNet Park et al. (2020). 0.982 0.993 0.969 0.981 –

VGG-16 Park et al. (2020). 0.932 0.979 0.877 0.925 –
ResNet-50 Park et al. (2020). 0.969 0.979 0.955 0.967 –
ANN Sayad et al. (2019) 0.983 0.970 0.980 0.970 –
CNN Wang et al. (2019) 0.907 – – – –
CNN Muhammad et al. (2018) 0.943 – – 0.890 –
CNN Li et al. (2019) 0.992 – – – –
Inception-V3 Alves et al. (2019) 0.948 0.997 0.988 0.992 –
Inception-V3 Sousa et al. (2020) 0.936 0.941 0.931 – –

specialized challenge of predicting the spread of fire in a short term to use this type of data in future studies as well as to extend the model by
period. integrating other land features. It is also crucial to consider an
Moreover, our model distinguishes itself from existing research improvement of the current model by focusing on the heterogeneity of
implementations by incorporating novel features outlined in Table 1, our dataset, especially for cases where there is a very small spread over
including wind direction, relative humidity, sea level pressure, and the 24h scale. For this, we aim to add more vegetation data that can help
precipitation. Furthermore, we represent these features as a time series predict the spread direction.
problem, providing a comprehensive depiction of their values for each In this research study, for each pair of perimeters, eleven maps are re-
hour over a 24-h period for every designated area of interest. This has a projected in the CRS of the satellite image with an UTM projection of 30
direct impact on the prediction accuracy of our model. On the one hand, m. The features that have been projected onto this resolution are DEM,
having an enhanced feature set, by incorporating wind direction, rela­ satellite bands 2–7, slope, aspect, initial and final burn maps. As for
tive humidity, and precipitation as additional features, our model cap­ weather data, we considered numerical representation for each area of
tures important meteorological factors that influence fire behavior. interest including a 24-h progression of climatic conditions. Alterna­
These features provide valuable contextual information that can tively, it might be interesting to use the spatial distribution of climatic
improve the accuracy of burned area predictions. In addition to that, this conditions as the same as other representative features. While we
enables the model to consider the temporal dynamics of meteorological already looked into obtaining these data from the High-Resolution
conditions. Wildfire behavior is highly influenced by the evolution of Rapid Refresh (HRRR) climatic data, using the HRRR-B library, this
weather conditions over time. By including hourly data over a 24-h method has two drawbacks, which led us to prioritize downloading
scale, our model can capture short-term variations and trends, allow­ numerical datasets. First, the dataset format requires that the download
ing for a more comprehensive understanding of fire dynamics. On the of atmospheric data, for a specified time, is for the entire United States.
other hand, the inclusion of these features on an hourly basis for specific This represents massive amounts of information that would make
areas of interest provides a spatial-temporal context to the model. downloading extremely time consuming for all area of interests that we
Wildfires can exhibit spatial heterogeneity, with different regions are working on. Moreover, with this technique, we would then have a
experiencing varying weather patterns. By incorporating localized map representing each meteorological attribute, for the 24 h following
meteorological data, our model can account for the spatial variations in the fire ignition. Although this is clearly of interest for assessing the
weather conditions and their impact on fire progression. sensitivity of the simulation results, the computational cost of processing
This comparative analysis is based on the latest recent research. In this large amount of data would be prohibitive.
terms of image recognition, fire and smoke detection models have In addition, our study represents a promising basis for the existing
generally achieved high performances using data augmentation tech­ literature to advance more research studies on fire spread prediction
niques as described in Table 3. DenseNet Li et al. (2019) and a CNN Park modelling, notably to study wildfire spread progression shortly after
et al. (2020) achieved accuracies of 98% and 99.2% respectively. In the their ignition. As discussed in the related work section, there have been
context of fire spread prediction, FU-NetcastV2 exceeded recent relatively few studies that include temperature, topography, and satel­
research studies. lite data to map predicted burned areas using a 24-h time scale. Fire or
In summary, our study overcomes the limitations of previous smoke detection systems are being given increased attention. Other deep
research by using a larger and more diverse dataset, evaluating perfor­ neural networks (Resnet and VGG) were also trained and tested, how­
mance using multiple metrics, and incorporating novel features repre­ ever due to their subpar performance, we have decided not to include
sented as a time series. These advancements enhance the reliability and them in Table 3.
effectiveness of our model for wildfire burned area prediction, paving The experimental results show that using meteorological data along
the way for more accurate and informed decision-making in wildfire with satellite and elevation features, FU-NetcastV2 can predict the size
management and mitigation efforts. of wildfires, providing a solid foundation for forest wildfire scale pre­
diction in the Rocky Mountains. As presented in the results section, our
5. Discussion model can be compared primarily to Radke, D. et al. Radke et al. (2019)
research, where the researchers implemented a CNN architecture to
The field of forest wildfire prediction is still open for further inves­ predict a future fire spread, based on atmospheric, vegetation and sat­
tigation and improvement. Its models could use additional data, such as ellite maps. They concatenated their eight atmospheric data points, for
firefighting intervention details, date and time of interventions, and the specific region, to the CNN output tensor, which is then used as an
vegetation index, which are not taken into account in this study. We aim input to a dense layer with a Sigmoid activation function, mapping to a

11
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

single output value for each pixel. Using 9 fires from the GeoMAC Ahmad, F., Goparaju, L., Qayum, A., 2018. Himalayan forest fire characterization in
relation to topography, socio-economy and meteorology parameters in Arunachal
dataset and data augmentation to increase their training set by a factor
Pradesh, India. Spatial Information Research 26, 305–315.
of 10, their model achieved an average accuracy of 87.7%. Akhloufi, M.A., Couturier, A., Castro, N.A., 2021. Unmanned aerial vehicles for wildland
We view our research as a valuable framework for looking into the fires: sensing, perception, cooperation and assistance. Drones 5, 15.
key fire parameters that affect fire spreading behavior 24 h after the Al-Kahlout, M.M., Ghaly, A.M.A., Mudawah, D.Z., Abu-Naser, S.S., 2020. Neural network
approach to predict forest fires using meteorological data. Int. J. Appl. Eng. Res. 4.
ignition. The gathered knowledge could be beneficial in locations other Alkhatib, A., Szmyt, J., 2017. Forest fire monitoring. In: Forest Fire. IntechOpen.
than the Rocky Mountains and the early data could assist and inform Alves, J., Soares, C., Torres, J.M., Sobral, P., Moreira, R.S., 2019. Automatic forest fire
firefighters letting them respond faster to major events. detection based on a machine learning and image analysis pipeline. In: World
Conference on Information Systems and Technologies. Springer, pp. 240–251.
de Bem, P.P., de Carvalho Júnior, O.A., Matricardi, E.A.T., Guimarães, R.F., Gomes, R.A.
6. Conclusion T., 2019. Predicting wildfire vulnerability using logistic regression and artificial
neural networks: a case study in Brazil’s federal district. Int. J. Wildland Fire 28,
35–45.
In the present paper we introduced FU-NetcastV2, a wildfire burned Bergado, J.R., Persello, C., Reinke, K., Stein, A., 2021. Predicting wildfire burns from big
area mapping model based on U-Net. Using consecutive GeoMac fire geodata using deep learning. Saf. Sci. 140, 105276.
Bjanes, A., De La Fuente, R., Mena, P., 2021. A deep learning ensemble model for wildfire
perimeters from the Rocky Mountains, satellite maps, DEM, the aspect,
susceptibility mapping. Ecol. Inf. 65, 101397.
the slope, and meteorological data, this model can forecast the burnt Boer, M.M., de Dios, V.R., Bradstock, R.A., 2020. Unprecedented burn area of australian
area after a 24-h period with an accuracy of 94.6% and an AUC of mega forest fires. Nat. Clim. Change 10, 171–172.
97.7%.The suggested algorithm could help fire fighters locate potential Boser, B.E., Guyon, I.M., Vapnik, V.N., 1992. A training algorithm for optimal margin
classifiers. In: Proceedings of the Fifth Annual Workshop on Computational Learning
burnt areas and therefore help them prevent the fire’s spread. A com­ Theory, pp. 144–152.
parable accuracy of 92.73% was attained with only 120 data points in Calp, M.H., Kose, U., 2020. Estimation of burned areas in forest fires using artificial
the previous version of the model Khennou et al. (2021). We do not neural networks. Ingeniería Solidaria 16, 1–22.
Cao, Y., Yang, F., Tang, Q., Lu, X., 2019. An attention enhanced bidirectional lstm for
believe that adding more data will considerably improve the results. early forest fire smoke recognition. IEEE Access 7, 154732–154742.
However, the inclusion of new features, such as the Normalized Differ­ Carter, J.V., Pan, J., Rai, S.N., Galandiuk, S.. ROC-ing along: Evaluation and
ence Vegetation Index (NDVI), fire extinguishing operations and in­ interpretation of receiver operating characteristic curves vol. 159, 1638–1645. URL:
https://www.sciencedirect.com/science/article/pii/S0039606016000660, doi:
terventions, may have a substantial influence on the fire behaviour. 10.1016/j.surg.2015.12.029..
This research advances the field of forest wildfire prediction by Crossing, V.. Weather data API | visual crossing. https://www.visualcrossing.com/we
taking into account other parameters such as terrain topography, ather-api. (Accessed 15 February 2021).
Dampage, U., Bandaranayake, L., Wanasinghe, R., Kottahachchi, K., Jayasanka, B., Forest
meteorological conditions, satellite images and elevation data. fire detection system using wireless sensor networks and machine learning vol. 12,
While our model can definitely be adapted to other regions, one of 46. URL: https://www.nature.com/articles/s41598-021-03882-9, doi:10.1038/
the approaches that may be relevant to consider beforehand, will be to s41598-021-03882-9. number: 1 Publisher: Nature Publishing Group..
Data.Gov. Datasets - CKAN. https://catalog.data.gov/dataset. (Accessed 20 January
emphasize on comparing feature importance methods and highlight
2021).
how satellite bands (6–7) impact the final burn. The results could help us Eftekharian, E., Ghodrat, M., He, Y., Ong, R.H., Kwok, K.C., Zhao, M., Samali, B., 2019.
reduce the dimensionality and analyze the computation time. Investigation of terrain slope effects on wind enhancement by a line source fire. Case
Stud. Therm. Eng. 14, 100467.
Elia, M., D’Este, M., Ascoli, D., Giannico, V., Spano, G., Ganga, A., Colangelo, G.,
CRediT authorship contribution statement Lafortezza, R., Sanesi, G., 2020. Estimating the probability of wildfire occurrence in
mediterranean landscapes using artificial neural networks. Environ. Impact Assess.
Rev. 85, 106474.
Fadoua Khennou: Conceptualization of this study, Methodology, Estes, B.L., Knapp, E.E., Skinner, C.N., Miller, J.D., Preisler, H.K., 2017. Factors
implementation. Moulay A. Akhloufi: Conceptualization of this study, influencing fire severity under moderate burning conditions in the klamath
Methodology, implementation. mountains, northern california, USA. Ecosphere 8, e01794.
Farhani, G., Sica, R.J., Daley, M.J., 2021. Classification of lidar measurements using
supervised and unsupervised machine learning methods. Atmos. Meas. Tech. 14,
Funding 391–402.
Fernández-Álvarez, M., Armesto, J., Picos, J., 2019. Lidar-based wildfire prevention in
wui: the automatic detection, measurement and evaluation of forest fuels. Forests 10,
This research was enabled in part by support provided by the Natural 148.
Sciences and Engineering Research Council of Canada (NSERC), funding Frizzi, S., Bouchouicha, M., Ginoux, J.M., Moreau, E., Sayadi, M., Convolutional neural
reference number RGPIN-2018-06233. network for smoke and fire semantic segmentation vol. 15, 634–647. URL: https
://onlinelibrary.wiley.com/doi/abs/10.1049/ipr2.12046, doi:10.1049/ipr2.12046..
Ganapathi Subramanian, S., Crowley, M., 2018. Using spatial reinforcement learning to
Acknowledgement build forest wildfire dynamics models from satellite images. Frontiers in ICT 5, 6.
Ghali, R., Akhloufi, M.A., Jmal, M., Mseddi, W.S., Attia, R., . Forest fires segmentation
using deep convolutional neural networks, in: 2021 IEEE International Conference
The authors would like to thank Jean-François Laplante for proof­ on Systems, Man, and Cybernetics (SMC), pp. 2109–2114. doi:10.1109/
reading this paper. SMC52423.2021.9658905. ISSN: 2577-1655..
Giglio, L., Justice, C., Boschetti, L., Roy, D., 2015. Mcd64a1 Modis. Terra+ Aqua Burned
Area Monthly L3 Global 500m SIN Grid V006 MCD64A1. https://doi.org/10.5067/
MODIS/MCD64A1.006.
Declaration of competing interest Güngöroğlu, C., 2018. Forest fire studies on fire behaviour: key topics and their
importance. In: Proceedings of the 14th International Combustion Symposium
(INCOS2018). Karabük, Turkey, pp. 25–27.
The authors declare that they have no known competing financial
Guo, F., Wang, G., Su, Z., Liang, H., Wang, W., Lin, F., Liu, A., 2016. What drives forest
interests or personal relationships that could have appeared to influence fire in fujian, china? evidence from logistic regression and random forests. Int. J.
the work reported in this paper. Wildland Fire 25, 505–519.
Guo, Y., Li, Z., Chen, E., Zhang, X., Zhao, L., Xu, E., Hou, Y., Liu, L., A deep fusion uNet
for mapping forests at tree species levels with multi-temporal high spatial resolution
Data availability satellite imagery vol. 13, 3613. URL: https://www.mdpi.com/2072-4292/13/1
8/3613, doi:10.3390/rs13183613. number: 18 Publisher: Multidisciplinary Digital
Data are public, details are in the paper Publishing Institute..
Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. Neural Comput. 9,
1735–1780.
References Hodges, J.L., Lattimer, B.Y., 2019. Wildland fire spread modeling using convolutional
neural networks. Fire Technol. 55, 2115–2142.
Holsinger, L., Parks, S.A., Miller, C., 2016. Weather, fuels, and topography impede
Abid, F., . A survey of machine learning algorithms based forest fires prediction and
wildland fire spread in western us landscapes. For. Ecol. Manag. 380, 59–69.
detection systems vol. 57, 559–590. URL: https://doi.org/10.1007/s10694-02
0-01056-z.

12
F. Khennou and M.A. Akhloufi Science of Remote Sensing 8 (2023) 100101

Hoover, K., Hanson, L.A.. Wildfire statistics. URL: https://apps.dtic.mil/sti/citations/ Park, M., Tran, D.Q., Jung, D., Park, S., et al., 2020. Wildfire-detection method using
AD1143321. section: Technical Reports. densenet and cyclegan data augmentation-based remote camera imagery. Rem. Sens.
Hopkinson, C., Chasmer, L., Young-Pow, C., Treitz, P., 2004. Assessing forest metrics 12, 3715.
with a ground-based scanning lidar. Can. J. For. Res. 34, 573–583. Radke, D., Hessler, A., Ellsworth, D., 2019. Firecast: Leveraging Deep Learning to Predict
Hristov, G., Raychev, J., Kinaneva, D., Zahariev, P., 2018. Emerging methods for early Wildfire Spread. IJCAI, pp. 4575–4581.
detection of forest fires using unmanned aerial vehicles and lorawan sensor Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: convolutional networks for
networks. In: 2018 28th EAEEIE Annual Conference (EAEEIE). IEEE, pp. 1–9. biomedical image segmentation. In: International Conference on Medical Image
Imada, A., 2014. A literature review: forest management with neural network and Computing and Computer-Assisted Intervention. Springer, pp. 234–241.
artificial intelligence. In: International Conference on Neural Networks and Artificial Rossa, C.G., Fernandes, P.M., 2018. Empirical modeling of fire spread rate in no-wind
Intelligence. Springer, pp. 9–21. and no-slope conditions. For. Sci. 64, 358–370.
Jain, P., Coogan, S.C., Subramanian, S.G., Crowley, M., Taylor, S., Flannigan, M.D., 2020. Sayad, Y.O., Mousannif, H., Al Moatassime, H., 2019. Predictive modeling of wildfires: a
A review of machine learning applications in wildfire science and management. new dataset and machine learning approach. Fire Saf. J. 104, 130–146.
Environ. Rev. 28, 478–505. Shmuel, A., Heifetz, E., 2022. Global wildfire susceptibility mapping based on machine
Jiao, Z., Zhang, Y., Xin, J., Mu, L., Yi, Y., Liu, H., Liu, D., 2019. A deep learning based learning models. Forests 13, 1050.
forest fire detection approach using uav and yolov3. In: 2019 1st International Simard, P.Y., Steinkraus, D., Platt, J.C., et al., 2003. Best Practices for Convolutional
Conference on Industrial Artificial Intelligence (IAI). IEEE, pp. 1–5. Neural Networks Applied to Visual Document Analysis. Icdar, Edinburgh.
Khennou, F., Ghaoui, J., Akhloufi, M.A., 2021. Forest fire spread prediction using deep Sofla, R.A.D., Alipour-Fard, T., Arefi, H., . Road Extraction from Satellite and Aerial
learning. In: Geospatial Informatics XI. International Society for Optics and Image Using SE-Unet 15, 014512. doi:10.1117/1.JRS.15.014512. publisher: SPIE..
Photonics, 117330I. Song, Y., Wang, Y., 2020. Global wildfire outlook forecast with neural networks. Rem.
Kinaneva, D., Hristov, G., Raychev, J., Zahariev, P., . Early forest fire detection using Sens. 12, 2246.
drones and artificial intelligence, in: 2019 42nd International Convention on Sousa, M.J., Moutinho, A., Almeida, M., 2020. Wildfire detection using transfer learning
Information and Communication Technology, Electronics and Microelectronics on augmented datasets. Expert Syst. Appl. 142, 112975.
(MIPRO), pp. 1060–1065. doi:10.23919/MIPRO.2019.8756696. ISSN: 2623-8764.. Subramanian, S.G., Crowley, M., 2017. Learning forest wildfire dynamics from satellite
Koutsias, N., Xanthopoulos, G., Founda, D., Xystrakis, F., Nioti, F., Pleniou, M., Mallinis, images using reinforcement learning. In: Conference on Reinforcement Learning and
G., Arianoutsou, M., Koutsias, N., Xanthopoulos, G., Founda, D., Xystrakis, F., Nioti, Decision Making. Ann Arbor MI.
F., Pleniou, M., Mallinis, G., Arianoutsou, M., . On the relationships between forest Survey, U.G., Landsat data access | u.s. geological survey. https://www.usgs.gov/landsat
fires and weather conditions in greece from long-term national observations -missions/landsat-data-access. (Accessed: 20 January 2022)..
(1894–2010) vol. 22, 493–507. URL: https://www.publish.csiro.au/wf/WF12003, Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D.,
doi:10.1071/WF12003. publisher: CSIRO PUBLISHING.. Vanhoucke, V., Rabinovich, A., 2015. Going deeper with convolutions. In:
Li, T., Zhao, E., Zhang, J., Hu, C., 2019. Detection of wildfire smoke images based on a Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
densely dilated convolutional network. Electronics 8, 1131. pp. 1–9.
Liang, H., Zhang, M., Wang, H., 2019. A neural network model for wildfire scale Tonini, M., D’Andrea, M., Biondi, G., Degli Esposti, S., Trucchia, A., Fiorucci, P., 2020.
prediction using meteorological factors. IEEE Access 7, 176746–176755. A machine learning-based approach for wildfire susceptibility mapping. the case
Ma, W., Feng, Z., Cheng, Z., Chen, S., Wang, F., 2020. Identifying forest fire driving study of the liguria region in Italy. Geosciences 10, 105.
factors and related impacts in china using random forest algorithm. Forests 11, 507. Wang, Y., Dang, L., Ren, J., 2019. Forest fire image recognition based on convolutional
Martell, D.L., Sun, H., . The impact of fire suppression, vegetation, and weather on the neural network. J. Algorithm Comput. Technol. 13, 1748302619887689.
area burned by lightning-caused forest fires in ontario vol. 38, 1547–1563. URL: Wang, Z., Yang, P., Liang, H., Zheng, C., Yin, J., Tian, Y., Cui, W., . Semantic
https://cdnsciencepub.com/doi/full/10.1139/X07-210, doi:10.1139/X07-210. segmentation and analysis on sensitive parameters of forest fire smoke using smoke-
publisher: NRC Research Press.. unet and landsat-8 imagery vol. 14, 45. URL: https://www.mdpi.com/2072-429
Mayr, M.J., Vanselow, K.A., Samimi, C., 2018. Fire regimes at the arid fringe: a 16-year 2/14/1/45, doi:10.3390/rs14010045. number: 1 Publisher: Multidisciplinary
remote sensing perspective (2000–2016) on the controls of fire activity in Namibia Digital Publishing Institute..
from spatial predictive models. Ecol. Indicat. 91, 324–337. Wildfire, R.T., . Rocky Mountain Research Station. https://www.fs.usda.gov/rmrs/gr
Minaee, S., Boykov, Y., Porikli, F., Plaza, A., Kehtarnavaz, N., Terzopoulos, D., Image oups/wire-wildfire-research. (Accessed: 23 May 2021)..
Segmentation Using Deep Learning: A Survey vol. 44, 3523–3542. doi:10.1109/ World, H.o. Wildfires. https://www.who.int/health-topics/wildfires#tab=tab_1.
TPAMI.2021.3059968. conference Name: IEEE Transactions on Pattern Analysis and (Accessed 23 May 2023).
Machine Intelligence.. Wulder, M., White, J., Alvarez, F., Han, T., Rogan, J., Hawkes, B., 2009. Characterizing
Mithal, V., Nayak, G., Khandelwal, A., Kumar, V., Nemani, R., Oza, N.C., 2018. Mapping boreal forest wildfire with multi-temporal landsat and lidar data. Rem. Sens.
burned areas in tropical forests using a novel machine learning framework. Rem. Environ. 113, 1540–1555.
Sens. 10, 69. Zhang, G., Wang, M., Liu, K., 2021. Deep neural networks for global wildfire
Muhammad, K., Ahmad, J., Baik, S.W., 2018. Early fire detection using convolutional susceptibility modelling. Ecol. Indicat. 127, 107735.
neural networks during surveillance for effective disaster management. Zhao, X., Lovreglio, R., Kuligowski, E., Nilsson, D., . Using artificial intelligence for safe
Neurocomputing 288, 30–42. and effective wildfire evacuations vol. 57, 483–485. URL: https://doi.org/10.1007
Naderpour, M., Rizeei, H.M., Ramezani, F., 2021. Forest fire risk prediction: a spatial /s10694-020-00979-x, doi:10.1007/s10694-020-00979-x..
deep neural network-based framework. Rem. Sens. 13, 2513. Zheng, H., Song, W., Wang, J., . Detrended fluctuation analysis of forest fires and related
Nasa’s, E.S.D.S., . Wildfires | Earthdata. https://earthdata.nasa.gov/learn/toolkits/ weather parameters vol. 387, 2091–2099. URL: https://www.sciencedirect.com/sci
wildfires. (Accessed: 26 May 2021).. ence/article/pii/S0378437107012204, doi:10.1016/j.physa.2007.11.020..

13