Abstract:
Settlement classifiers for multitemporal satellite
image analysis have to overcome numerous difficulties related
to across-date variances in viewing- and illumination geometry.
Shadow anisotropy is a prominent contributing factor in classifier
inaccuracy, so methods are introduced in this study to enable
minimum-supervision classifier design that mitigate the effects
of shadow profile differences. A segmentation-based shadow detector
is proposed that utilizes a panchromatic segment merging
algorithm with parameters that are robust against dynamic range
variances seen in multitemporal imagery. The proposed shadow
detector improves on the settlement classification accuracy
achieved by fixed threshold detection paired with shadow removal
in the presented case-study. The relationship between shadow
detection accuracy and settlement classification accuracy is investigated,
and it is shown that shadow removal produces greater
settlement accuracy improvements for across-date experiments
specifically.