Database Reference
In-Depth Information
ful parameters related to the event that can be discovered by analyzing
spatial-temporal data.
Time series change detection approaches can be broadly categorized
under parameter change detection [15, 37, 68], anomaly detection [10,
36, 21, 63], prediction based methods [24, 73, 44, 40], and segmentation
based approaches [34, 47, 46, 2]. Parameter change detection approaches
assume that the detected event will exhibit a change in a characteristic
parameter such as the mean or the variance of the data, and hence they
can be identified by monitoring changes in the distribution of this statis-
tic. They have been used in the past for mapping forest fires in Portugal
as an example [45]. Anomaly detection approaches find subsequences
that are unusual with the underlying hypothesis that observations devi-
ate from normal when an event occurs [62]. Anomaly detection based on
discord discovery finds subsequences that are significantly dissimilar to
all other subsequences of the time series [18, 43]. Prediction based ap-
proaches explicitly learn a model that predicts future observations based
on previous values, and deviation of the observed data from the model
prediction is used as an indicator of change [31, 51]. Segmentation based
approaches divide a time series into homogeneous parts (that can be ap-
proximated by a simpler generative process), where segments have high
intra-segment similarity but low inter-segment similarity [28, 74, 2].
In the spatio-temporal context, event detection has been studied in
the domain of sensor networks such as in [79, 77]. In the earth science
domain, there is limited work that utilizes both spatial and temporal
components effectively. Examples include [22] which characterize spatio-
temporal fire activity patterns using satellite imagery, where airborne
images are used for spatiotemporal change detection in forest cover [61],
and classification-based approaches [9, 16].
A series of algorithms for land cover change detection [50] have re-
cently been developed using predictive [51], segmentation [3, 4, 28] and
parameter change [8] approaches on vegetation based time series data.
These highly scalable algorithms overcome many of the limitations of
traditional approaches to land cover change detection including the sus-
ceptibility to noise and temporal variability. The algorithms have been
comparatively evaluated with state of the art land cover change detec-
tion techniques, and applied to global vegetation data (EVI) to detect a
variety of changes in the global ecosystem including those due to fires,
deforestation, insect damage, floods, hurricanes, conversion to agricul-
ture, urbanization [50, 29, 51, 28, 8, 4, 5, 2].
