Please use this identifier to cite or link to this item:
Title: Land Cover Classification Using Multi-Frequency SAR over Semi-Arid and Forested African Landscapes
Authors: Spies, Bernardus Francois
Supervisors: Balzter, Heiko
Award date: 21-Nov-2016
Presented at: University of Leicester
Abstract: The potential of using multi-frequency Synthetic Aperture Radar (SAR) for land cover classification is becoming a reality, with multiple SAR satellites utilising different frequencies currently in orbit and more missions planned for the future. This study looks at combining SAR frequencies from L-band (ALOS PALSAR), C-band (ENVISAT ASAR) and X-band (TerraSAR-X) to find the optimum combination of the SAR data for land cover classification of forested and semi-arid ecoregions in Africa. The study site for forested areas is in Cameroon and the semi-arid study site is in Tanzania. Data from both the wet and dry seasons are available. Random forest models, with different combinations of input variables, are compared. Models with the top 30 variables are chosen from the mean decrease accuracy and mean decrease Gini variable importance measures, and compared with the classification accuracies using support vector machines. Some of the findings are that L-band is the best single frequency for land cover classifications for both ecoregions, with X-band the best single frequency if only forested regions are considered. Texture measures lead to an increase of between 15-25% overall classification accuracy compared to using only backscatter coefficients. The recommended dual-frequency combination are LX-bands, although L-band data give overall classification accuracies very close to LX-bands. The use of images from LCX data only marginally improves the classification accuracy from LX-images and L-band images. The benefit from acquisition of all three frequencies would therefore rarely outweigh the cost of acquiring and processing data from all three frequencies. The transferability of the random forest models to an additional geographic site did not produce satisfactory results, however the transferability of the random forest models to additional season data did give satisfactory results. The Kullback-Leibler divergence class difference measure showed potential to give an indication of transferability of the models, although refinement remains necessary.
Type: Thesis
Level: Doctoral
Qualification: PhD
Rights: Copyright © the author. All rights reserved.
Appears in Collections:Leicester Theses
Theses, Dept. of Geography

Files in This Item:
File Description SizeFormat 
2016SPIESBFPhD.pdfThesis29.65 MBAdobe PDFView/Open

Items in LRA are protected by copyright, with all rights reserved, unless otherwise indicated.