Classification of remote sensing images using support vector machines

Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Pramod K. Varshney


Classification, Remote sensing images, Support vector machines

Subject Categories

Electrical and Computer Engineering | Engineering


Land cover information is essential for many diverse applications. Various natural resource management, planning, and monitoring programs depend on accurate information about the land cover in a region. Remotely sensed images are attractive sources for extracting land cover information, where an image classification algorithm is employed to retrieve land cover information. Existing classifiers have shown marked limitations. Recently, Support Vector Machines (SVMs) have been proposed as an alternative to produce land cover classification. The goal of this dissertation is to conduct an extensive study of SVMs for this task and to develop novel SVM based classification algorithms.

In this dissertation, we consider SVM based classification of both multispectral and hyperspectral data from remote sensing sensors. The construction of a classifier based on SVM algorithms (for multiclass classification problems) is investigated in detail via experiments with real data. Factors such as the choice of the multiclass method, the optimizer and the kernel function are examined in detail. The efficacy of feature extraction, as a pre-processing step to SVM classification to reduce the dimensionality of the data, is also examined. It is confirmed that feature extraction degrades the performance of SVM based classifiers. Training sample size is an important consideration for the successful implementation of any classifier; therefore, the sensitivity of SVM based classifiers with respect to training sample size is evaluated. It is shown that these classifiers perform quite well with a small training sample size. The performance of the SVM is compared with other competitive and well known classifiers for both multispectral and hyperspectral data. SVM based classifiers exhibit superior performance. A novel SVM based approach for multisource classification technique is introduced. It is shown that use of ancillary data and information fusion enhances classification performance. Finally, two new approaches that employ SVMs for soft classification of images dominated by mixed pixels are proposed and their performance evaluated. They are shown to perform better than several existing soft classification algorithms.


Surface provides description only. Full text is available to ProQuest subscribers. Ask your Librarian for assistance.