This study explored the utility of an object-based image classification approach for mapping land cover in a heterogeneous coastal zone using WorldView-2 imagery. Two relatively modern and robust supervised machine learning algorithms i.e. random forest (RF) and support vector machines (SVM) were also compared. Image segmentation was performed, and ten broad land cover classes were identified. Subsequently, we assessed the performance of an object based image classification and the selected machine learning algorithms in mapping the land cover classes. The validation of the thematic land cover maps derived from RF and SVM were assessed using an independent test dataset generated from field work data and aerial photography interpretation. Results showed that both the machine learning classifiers in combination with the object-based approach are useful in mapping land cover in heterogeneous coastal areas. However, SVM achieved the best overall accuracy (93.79%) and kappa statistic (0.93) while RF produced an overall accuracy of 86.94% and kappa value of 0.85. Overall, the study underlined the utility of combining an objectbased image classification with machine learning classifiers for mapping land-cover in heterogeneous coastal areas – a previously challenging task with broad band satellite sensors and traditional pixel-based image classification approaches.

Date Published: 25 October 2012
PDF: 10 pages
Proc. SPIE 8538, Earth Resources and Environmental Remote Sensing/GIS Applications III, 85380P (25 October 2012); doi: 10.1117/12.978996

Published in SPIE Proceedings Vol. 8538:
Earth Resources and Environmental Remote Sensing/GIS Applications III
Shahid Habib; Ulrich Michel; Daniel L. Civco; David Messinger; Antonino Maltese; Manfred Ehlers; Karsten Schulz; Konstantinos G. Nikolakopoulos, Editor(s)