Comparison of Supervised Algorithms on DIWATA-I Microsatellite Space Bourne Multispectral Imagery

Abstract

The Phl-Microsat program in the Philippines was initiated for capacity building and with the end goal of having a source of remotely-sensed data for local planning, disaster risk mitigation, and resource management for the country. To increase its benefits, an established process to effectively utilize these images such as image classification is needed. This study aims to determine the most appropriate supervised algorithm for image classification among a set of classifiers that will yield the best results for DIWATA-I Spaceborne Multispectral Images (SMI). SMI is an optical payload, with 80m resolution, and a multiwavelength selection at 10nm width at 1nm steps. Three study sites within the Philippines were selected to test the classifiers - Camarines Sur, Ilocos Norte, and Oriental Mindoro. Spectral reflectance values were then derived from atmospheric calibrations of the images. These images were then classified using six supervised classifiers and were post-processed using Majority Analysis. Accuracy is assessed by comparing the overall accuracy, kappa coefficient, producer’s accuracy and user’s accuracy extracted from the confusion matrix. From the results, Support Vector Machine and Maximum Likelihood classifiers produced the most desirable and most consistent results.