Classification of Images Using Support Vector Machines

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.

💡 Research Summary

This paper investigates two widely used multiclass extensions of Support Vector Machines (SVM) – One‑Against‑One (1A1) and One‑Against‑All (1AA) – in the context of land‑cover mapping from remotely sensed imagery. SVMs are fundamentally binary classifiers that achieve high generalisation performance by projecting input data into a high‑dimensional feature space via kernel functions and then maximising the margin between classes. Because most remote‑sensing tasks involve more than two land‑cover categories, binary SVMs must be adapted. The 1AA approach builds N binary classifiers for an N‑class problem, each separating one class from the rest; it is simple and computationally cheap but can suffer from class‑imbalance and often produces unclassified or “mixed” pixels when decision boundaries overlap. The 1A1 approach constructs a classifier for every pair of classes, yielding N(N‑1)/2 models; it typically yields cleaner class boundaries and fewer ambiguous pixels, at the cost of substantially higher computational demand.

The authors conducted an empirical comparison using a 2001 Landsat scene (rows 171, columns 60) covering the source of the River Nile in Jinja, Uganda. Six optical bands (1‑5, 7) were employed, and three land‑cover classes – water, vegetation, and built‑up – were defined. Preliminary data handling (scene extraction, training sample selection) was performed in IDRISI Andes, after which the data were exported to MATLAB 7 for SVM training and classification. Four kernel types were examined: linear, polynomial (degree 2), quadratic, and radial basis function (RBF). Each kernel was applied using both 1AA and 1A1 schemes, producing eight classification results.

Four evaluation criteria were used: (1) number of unclassified pixels, (2) number of mixed pixels (pixels assigned to more than one class), (3) overall accuracy measured by the Kappa statistic, and (4) statistical significance of the accuracy difference between the two schemes, assessed via a Z‑test (|Z| > 1.96 indicating significance at the 95 % confidence level).

Results show a consistent pattern: the 1AA scheme generates far more unclassified and mixed pixels than 1A1. For the linear kernel, 1A1 left only 16 unclassified pixels, whereas 1AA left 700; mixed pixels were absent in 1A1 but numbered 9 048 in 1AA. Similar trends appear for the polynomial, quadratic, and RBF kernels, although the RBF case produced no mixed pixels. Despite these differences in pixel‑level ambiguity, overall classification accuracies were very high for all configurations (Kappa ≥ 0.88). The linear kernel achieved Kappa = 1.00 (1A1) vs 0.95 (1AA); quadratic gave 0.88 vs 0.94; polynomial yielded 1.00 for both; RBF produced 0.97 vs 0.92. The Z‑values for all comparisons were below 1.96, indicating that the accuracy differences are not statistically significant at the 95 % level.

Visual inspection of the resulting land‑cover maps corroborates the quantitative findings. Maps derived from 1A1 appear smoother and more aesthetically pleasing, while 1AA maps contain conspicuous black spots (unclassified) and red speckles (mixed), which would require post‑processing (e.g., aggregation or manual correction) before analysis.

The authors conclude that, from an accuracy standpoint, 1A1 and 1AA are essentially equivalent for the dataset and classifiers examined. The choice between them should therefore be guided by practical considerations: computational resources, the degree of class imbalance, and the tolerance for ambiguous pixels. When computational capacity is limited or rapid prototyping is needed, 1AA offers a viable shortcut, provided that downstream processing can handle the higher rate of unclassified/mixed pixels. Conversely, when precise class boundaries are critical—such as in detailed urban planning or change‑detection studies—the extra computational cost of 1A1 is justified.

The study contributes an empirical benchmark for SVM multiclass strategies in remote sensing, highlighting that methodological selection can be driven by dataset characteristics and project requirements rather than inherent superiority of one scheme. Future work could explore hybrid schemes, adaptive weighting of votes in 1A1, or automated cleaning of 1AA‑produced ambiguous pixels, especially for higher‑resolution or hyperspectral imagery where the number of classes and data volume are larger.