IEEE Geoscience and Remote Sensing Magazine - June 2023 - 14

WE NEED TO FIGURE OUT
WAYS OF PRODUCING
REPRESENTATIVE ENOUGH
DATA SO THAT THE
PERFORMANCE OF A
SOLUTION CAN BE
VALIDATED BEFORE FLIGHT,
ESPECIALLY IN DATADRIVEN
APPROACHES.
Here, remote sensing, which captures detailed plant characteristics
by using, e.g., hyperspectral sensors, brings
important practical advantages of being noninvasive
and inherently scalable over larger areas. In the case of
ground-based methods, a hyperspectral camera is commonly
placed at a distance of about 30 cm above the test
object, with constant lighting conditions [25]. Since
various parts of the available spectrum are important
in different use cases, the raw hyperspectral data are often
preprocessed with band selection methods [26]. To
automate the analysis process of highly dimensional hyperspectral
data, an array of
machine learning methods
were introduced for monitoring
plant diseases [18]. Deep
learning techniques
constitute
the current research
focus in the field, with deep
convolutional neural networks
(CNNs) playing the leading
role [27], [28], [29]. On the
other hand, there are approaches
that strongly benefit
from handcrafted features
and vegetation indices that
may be extracted from satellite
MSIs [23].
The impact of global climate change is manifested by
drought in terrestrial ecosystems. Drought is one of the
main factors of environmental stress; therefore, estimating
the water content in vegetation is one of the critically
important practical challenges faced today [30].
Based on this information, it is possible to monitor the
physiology and health of forests and crops [31] by extracting
an array of quantifiable drought-related parameters.
Such parameters include, but are not limited to,
the canopy water content (reflecting water conditions at
the leaf level and information related to the structure of
the canopy, based on remote sensing data [32]), the leaf
equivalent water thickness (estimating the water content
per unit of a leaf's area [33]), and the live fuel moisture
content [34]. There are, indeed, in-field methods that
may allow us to quantify such parameters, but they are
not scalable; hence, they are infeasible for large areas of
interest [30].
DETECTION AND MONITORING OF FLOODS
Between 1995 and 2015, flooding affected more than
2.2 billion people, representing 53% of all people affected
by weather-related disasters. Detecting and monitoring
floods in vulnerable areas is extremely important in emergency
activities to maintain the safety of the population
and diversity of the underlying ecosystems [35]. Today, the
most important source of information about floods is satellite
data; both MSI and synthetic aperture radar (SAR) data
are used to detect and determine floods' extent [35], [36]. A
14
key element in satellite-based flood monitoring is the need
for rapid disaster response.
Detection and assessment of floods is difficult due to
the environmental impact of the water present in affected
areas [37]. SAR has great potential to monitor flood situations
in near real time, due to its ability to monitor Earth's
surface in all weather conditions [36]. Such data have been
successfully exploited in various deep learning-powered
solutions toward the detection of floods [36], [38]. Here,
temporal information captured within time series data
plays a key role in quantifying and understanding the
evolution of flood maps. Detection of flooding areas is
also performed based on MSIs, commonly with the use
of the Normalized Difference Water Index [35]. Interestingly,
data-driven solutions may benefit from the public
WorldFloods database, which currently contains 422
flood maps [35].
DETECTION OF FIRE, VOLCANIC
ERUPTIONS, AND ASH CLOUDS
Fires are an example of a natural threat that destroys natural
resources and causes extensive socioeconomic damage.
The nature of a fire is primarily influenced by the type,
flammability, and amount of fuel present in an area. Climate
change and natural conditions, such as topography
and wind, are also important factors here. Historically, the
estimation of a fire area was carried out using field methods
based on GPS data. However, only the perimeter of
the area can be determined, and this approach is limited
by the difficulty in reaching an area covered by an active
fire and the risk of unevenness of burning areas [39]. The
nature of a fire also changes over time, due to the fire's
effect of increasing the volume of charred vegetation and
changing the temperature and level of humidity. Importantly,
the decreasing amount of chlorophyll as a result of
combustion causes changes in the characteristics of the
spectral signature of acquired MSIs/HSIs [40]. This information
allows us to assess whether an area is experiencing
an active fire (because there are changes over time), the
area is burned (because the chlorophyll content is low),
and there is a risk of the active area redeveloping (partial
fuel burnout) [41].
The monitoring and prevention of fires in Europe is carried
out by the European Forest Fire Information System,
which is part of the Copernicus EO program, in which the
monitoring process exploits the 13-band MSI data captured
by the Sentinel-2 mission. The key source of information
here is the red-edge band, which is one of the best descriptors
of chlorophyll, whereas the assessment of a fire condition
is most often based on vegetation indices [40]. Such
analysis can be, however, automated using a variety of machine
learning approaches operating on MSIs and HSIs [42],
[43]. Accurate detection of active areas of fire, obtained in
a safe way, is an important advantage over methods based
only on the perimeter of an area, including both active and
burned parts [40]. Moreover, the application of remote
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2023
IEEE Geoscience and Remote Sensing Magazine - June 2023

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