Feature selection for sky image classification based on self adaptive ant colony system algorithm

Abstract

Statistical-based feature extraction has been typically used to purpose obtaining the important features from the sky image for cloud classification. These features come up with many kinds of noise, redundant and irrelevant features which can influence the classification accuracy and be time consuming. Thus, this paper proposed a new feature selection algorithm to distinguish significant features from the extracted features using an ant colony system (ACS). The informative features are extracted from the sky images using a Gaussian smoothness standard deviation, and then represented in a directed graph. In feature selection phase, the self-adaptive ACS (SAACS) algorithm has been improved by enhancing the exploration mechanism to select only the significant features. Support vector machine, kernel support vector machine, multilayer perceptron, random forest, k-nearest neighbor, and decision tree were used to evaluate the algorithms. Four datasets are used to test the proposed model: Kiel, Singapore whole-sky imaging categories, MGC Diagnostics Corporation, and greatest common divisor. The SAACS algorithm is compared with six bio-inspired benchmark feature selection algorithms. The SAACS algorithm achieved classification accuracy of 95.64% that is superior to all the benchmark feature selection algorithms. Additionally, the Friedman test and Mann-Whitney U test are employed to statistically evaluate the efficiency of the proposed algorithms.