IEEE Geoscience and Remote Sensing Magazine - September 2023 - 107

the few other datasets about wildfires [11], [12], our dataset
covers a larger area and spans more years.
Ground truth masks for the task of binary image segmentation
were generated starting from the public vector
data provided by California's Department of Forestry and
Fire Protection [13] and rasterized. Satellite acquisitions,
i.e., the raw input data, were instead collected from the
Sentinel-2 L2A mission through Copernicus Open Access
Hub. More precisely, we collected and released both
prefire and postfire information associated with the same
area of interest.
The contributions of this article can be summarized as
follows:
◗ A novel image segmentation dataset was tailored to burned
area delineation consisting of Sentinel-2 pre- and postfire
acquisitions. We provide more samples than existing datasets
to facilitate the training of (large) deep learning models.
◗ Three different baselines were evaluated on the proposed
dataset: one consisting of the evaluation of several
burned area indexes and Otsu's automatic thresholding
method [14], one based on the SegFormer model
[15], and one based on the U-Net model [16].
The article is structured as follows. The " Related Works "
section introduces the related works, the " Dataset " section
introduces the collected dataset and the preprocessing steps
performed, and the " Tasks " and " Experiments " sections
formally introduce the tasks and the experimental settings
and results. Finally, the final section concludes the article.
RELATED WORKS
Before the development of deep learning-based methodologies,
domain experts based their analyses on satellite
imagery leveraging spectral index computation and evaluation.
Considering the synthetic aperture radar context,
thresholding-based techniques have been adopted to distinguish
between flooded and unflooded areas [17]. Different
analyses have been performed on various tasks concerning
several spectral indexes, such as in cloud detection
(cloud mask) [18], water presence (water pixels and the normalized
difference water index) [19], [20], and vegetation
analysis (the normalized difference vegetation index) [21].
Considering the burned area delineation problem, domain
experts have developed several indexes: the Normalized
Burn Ratio (NBR), NBR2, Burn Area Index (BAI), and
BAI for Sentinel-2 (BAIS2) [19]. They are computed using
different spectral bands to generate an index highlighting
the affected areas of interest. Such techniques are often
coupled with thresholding methodologies: either fixed or
manually calibrated threshold values are chosen [22], or automatic
thresholding algorithms are used [23]. Additional
studies evaluate index-based techniques with additional
in situ information, namely, the Composite Burned Area
Index, which, indeed, provides insightful information but
does not represent a scalable solution because in situ data
are incredibly costly to collect. Furthermore, studies confirmed
that finding a unique threshold that is region and
SEPTEMBER 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
vegetation independent is difficult [24]. These methods
assume that burned and unburned areas are linearly separable,
which is usually untrue.
More recently, researchers started adopting supervised
learning techniques to solve several tasks in computer vision
and EO. More precisely, convolutional neural network
(CNN)-based models proved their effectiveness in image
classification and segmentation tasks, achieving state-ofthe-art
performances compared to index-based methodologies
[25], [26]. Deep models proved their effectiveness in
similar tasks covering wildfire detection [27] and spreading
[28], too. The main drawback is the need for a significant
amount of labeled data, possibly
covering heterogeneous
regions with different morphological
characteristics, to
learn better representations.
Over the years, many of the
proposed frameworks have
limited their analyses to a few
samples collected from a limited
number of countries or
locations [29]. In a few cases,
larger datasets were adopted
to tackle the semantic segmentation
problem without
disclosing the dataset [27].
In the EO domain, different public datasets are availIT
IS WELL KNOWN THAT
MORE CURATED DATA
PROVIDE BETTER MACHINE
LEARNING MODELS, AND
THE DATASET PROVIDES
MANY READY-TO-USE
SAMPLES WITHOUT
LEVERAGING OTHER
SOURCES.
able to the research community tackling different problems,
such as flood delineation [17], [30], deforestation
[31], wild area monitoring [32], sustainable development
goal monitoring [33], and crop classification and segmentation
[34], but, to the best of our knowledge, only
two public datasets are available for the burned area delineation
problem covering some countries in Europe
[11] and Indonesia [12]. Our dataset collects more data
than these, considering more wildfires and a larger area.
It comprises pre- and postfire Sentinel-2 L2A data about
California forest fires.
Table 1 shows a comparison among the three datasets.
The proposed dataset consists of the highest number
of considered wildfires (340), globally covering the largest
amount of burned areas (28 million pixels covering
11,000 km2) and a higher total covered surface (450,000 km2).
Figure 1(a) shows the covered areas. Even though the
proposed dataset has the greatest amount of burned surface,
it achieves the lowest percentage of burned area
compared to the others. However, the CaBuAr dataset
provides the highest number of training samples in supervised
(supervised learning in binary segmentation and the
highest amount of burned areas) and unsupervised cases
(self-supervised learning and the highest area covered). It
is well known that more curated data provide better machine
learning models, and the dataset provides many
ready-to-use samples without leveraging other sources.
Images are larger in terms of pixels (5,490) and disclosed
107
IEEE Geoscience and Remote Sensing Magazine - September 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2023
