Database Reference
In-Depth Information
In the remainder of the section, we present two illustrative examples
of change detection using remote sensing datasets - (i) land cover change
monitoring, and (ii) identifying ocean eddy dynamics.
4.1 Illustrative Application: Monitoring
Changes in Land Cover
Detecting meaningful events from global-scale earth science datasets
poses several unique challenges which are yet to be addressed by tradi-
tional event detection techniques in the domain. Approaches that utilize
only the spatial information using image snapshots disregard the rich
temporal context of the data and are significantly impacted by temporal
variability and noise. On the other hand, time series change detection
methods do not utilize spatial information and thus are limited in their
global applicability. Earth science datasets show varying characteristics
in different climatic conditions and land cover types and are thus spa-
tially heterogeneous. At the same time, locations that are spatially close
to each other exhibit similar temporal variability due to spatial autocor-
relation. This property influences the creation of a background model
of unchanged locations in varying spatial regions. Furthermore, earth
science datasets show similar values during the same seasonal duration
across different years, due to the annual cycles of the earth. Addition-
ally, similar event occurrences show spatial autocorrelation, which can
be leveraged for improved event detection of spatially coherent events
with low-intensity of change at varying degrees of representation.
Research on land cover change detection falls in three major cate-
gories namely spatial change detection, temporal change detection and
spatio-temporal change detection. In the spatial domain, change detec-
tion has been framed as a classification problem and various approaches
such as Markov random field (MRF) based methods [39], Gibbs-MRF
[65] and Gaussian-MRF [80] have been developed. In addition, image
comparison-based approaches (that compare snapshots of a region from
different time steps usually separated by multiple years) have been de-
veloped in the earth science community and used to identify events such
as urbanization, forest disturbances and agriculture related changes [20].
A major limitation of these methods is that they are inherently region-
specific as high-quality training data is expensive to generate, and ac-
curacy is poor if a classifier learned from the training samples of one
geographical region is used for classifying test samples from another re-
gion. Furthermore, these methods fail to exploit the rich information
in the temporal context. In particular, these methods are unable to
identify the exact change duration, the rate of change and other use-
