Classification using multi-sensor imagery
Student name: Mr Ashish Kumar Agarwalla
Guide: Dr Vinay Shankar Prasad Sinha
Year of completion: 2016
Host Organisation: Jawaharlal Nehru University, New Delhi
Supervisor (Host Organisation): Prof Sonajharia Minz
Abstract: Urban environments are entitled to have high spatial and spectral diversity. Remotely sensed
imagery from a single source, accounts for sensor specific information only. Use of data from
multiple sources and similar temporal resolution could account for necessary complementary
information missing from the individual source and help improve classification accuracy. The
following project, aimed at developing an unsupervised classification algorithmfor improved
classification of remotely sensed imagery, for data collected from multiple sources as compared
to a single source and no training involvement. This was done by clustering a random mask of
100 X 100 pixels from multispectral data from Landsat TM and evaluation of cluster quality to
find the number of naturally occurring clusters. Similarly, four more random masks were used to
predict 12 naturally occurring clusters in the data. This was followed by clustering the entire MS
data using k-means algorithm and evaluation of the resulting cluster quality using silhouette
coefficient to identify loosely classified pixels at mean silhouette value (threshold). Hyperspectral
data from Hyperion was used to extract data for only the loosely classified pixels
identified above and was clustered using the k-means algorithm. Finally, the clustered output
from HS data was fused with good quality clusters (clusters with silhouette coefficient above the
mean) from the MS imagery to produce final classified maps.
In the fused imagery,the overall Classification accuracy and Kappa Statistics increased
significantly as compared to the MS imagery. Further attempts were made to lower the threshold
to half, but this lowered the classification accuracy to78.57 % and Kappa Coefficient to
0.7.Thus, Silhouette coefficient proved to be a good evaluator of cluster quality and predictor of
naturally occurring clusters. Further, it can be predicted that fusion of data from multiple
sources yields better classification results.
Keywords: Unsupervised Classification, Clustering, k-means, Silhouette Coefficient, Data
fusion.
