[Paper Review] Classification of Images Using Support Vector Machines

Anthony Gidudu, Greg Hulley+1 more|ArXiv.org|Sep 25, 2007

Remote-Sensing Image Classification11 references38 citations

TL;DR

This paper evaluates One-Against-One (1A1) and One-Against-All (1AA) strategies for applying Support Vector Machines (SVMs) to multi-class land cover image classification. While 1AA tends to produce more unclassified and mixed pixels, both methods yield comparable classification accuracy, leading the authors to conclude that the choice between them is largely dataset-dependent and a matter of personal preference.

ABSTRACT

Support Vector Machines (SVMs) are a relatively new supervised classification technique to the land cover mapping community. They have their roots in Statistical Learning Theory and have gained prominence because they are robust, accurate and are effective even when using a small training sample. By their nature SVMs are essentially binary classifiers, however, they can be adopted to handle the multiple classification tasks common in remote sensing studies. The two approaches commonly used are the One-Against-One (1A1) and One-Against-All (1AA) techniques. In this paper, these approaches are evaluated in as far as their impact and implication for land cover mapping. The main finding from this research is that whereas the 1AA technique is more predisposed to yielding unclassified and mixed pixels, the resulting classification accuracy is not significantly different from 1A1 approach. It is the authors conclusions that ultimately the choice of technique adopted boils down to personal preference and the uniqueness of the dataset at hand.

Motivation & Objective

• To evaluate the effectiveness of One-Against-One (1A1) and One-Against-All (1AA) strategies in multi-class image classification using SVMs.
• To assess the impact of these strategies on classification accuracy and pixel output quality in land cover mapping.
• To determine whether one approach consistently outperforms the other in remote sensing applications.
• To provide guidance on selecting between 1A1 and 1AA based on dataset characteristics and user preferences.

Proposed method

• The study applies SVMs to remote sensing image data, leveraging their robustness and effectiveness with small training samples.
• Two multi-class classification strategies are implemented: One-Against-One (1A1), which trains a binary classifier for each pair of classes.
• One-Against-All (1AA), which trains a binary classifier for each class against all others.
• Classification results are compared across both strategies using standard accuracy metrics.
• The evaluation focuses on classification accuracy, the number of unclassified pixels, and the presence of mixed pixels.
• The analysis is conducted on real remote sensing datasets relevant to land cover mapping.

Experimental results

Research questions

• RQ1 How do the 1A1 and 1AA SVM strategies compare in terms of classification accuracy for land cover mapping?
• RQ2 What is the impact of each strategy on the number of unclassified and mixed pixels?
• RQ3 Does one strategy consistently produce better results across different datasets?
• RQ4 How do dataset-specific characteristics influence the choice between 1A1 and 1AA?

Key findings

• The One-Against-All (1AA) approach produces a higher number of unclassified and mixed pixels compared to the One-Against-One (1A1) approach.
• Despite differences in pixel output quality, the overall classification accuracy between 1A1 and 1AA is not significantly different.
• The 1AA method shows a tendency to misclassify or leave out certain pixels due to its binary classification structure.
• The study finds no statistically significant performance advantage for either 1A1 or 1AA in terms of accuracy.
• The final choice between 1A1 and 1AA is determined more by dataset-specific traits and user preference than by objective performance metrics.