On the performance analysis of classifier fusion for land cover classification

We investigate the performance evaluation of merging (fusing) the classification capabilities of classifiers for the land use analysis. For the fusion approach, we select the parametric and non-parametric classifiers. The set used includes: Bayesian Network, Multi-layer Perceptron (MLP), Support Vector Machines (SVM) and Random Forest. These classifiers are selected based on their good over-all performance for the land use analysis and in general for other classification tasks. We evaluate the concept on both the high and low resolution multispectral satellite imagery. The performance of the approach is evaluated using F-score, computation time and accuracy. Based on the experimental evaluation, we advocate the use of classifier fusion for the low resolution satellite imagery. While for high resolution satellite imagery, the fusion shows slight improvement in performance.