Weather classification from still images using ensemble method

Abstract

In the field of computer vision, multi-class outdoor weather classification is a difficult task to perform due to diversity and lack of distinct weather characteristic or features. This research proposed a novel framework for identifying different weather scenes from still images using heterogeneous ensemble methods. The approach was based on construction of unobstructed opaque cloud coverage (OCC) multi-class weather images; and the introduction of diversity concept called Selection Based on Accuracy Intuition and diversity (SAID) for the construction of stacked ensemble models. The stages involve the extraction of histogram of features from different weather scenes to determine their contribution to the overall performance of the experiment, training and validating the performance of the model. The blending and boosting of different weather features using stacked ensemble algorithms shows an average accuracy of over 90% in recognizing rainy still images and over 80% for sunny, sunrise and sunset still images. Similarly, the meta-learner of the stacked ensemble model performed better than the individual base learners of the model. The research presents academic and practitioners a new insight into diversity of heterogeneous stacked ensemble methods for solving the challenges of weather recognition from still images.