Cloud Classification in Sky Images using Deep Neural Networks

Abstract

In this research endeavor, a thorough analysis of a 10-minute sky video sequence was conducted. The study commenced by extracting frames at a consistent rate of 30 frames per second, resulting in a dataset of 1200 cloud images. Following frame extraction, color image segmentation was applied to identify distinct color regions within each frame.The primary goal was to estimate the percentage of cloudy pixels within each frame. To achieve this, three pre-trained Convolutional Neural Network (CNN) models namely - VGG16, MobileNetV2, and ResNet50 - were employed for cloud detection and pixel classification. This three-model approach contributed to a comprehensive assessment of cloud cover in the images. Cloudy pixel percentage was calculated using both area-based and pixel count-based approaches, adding depth to the analysis. This holistic approach provided a nuanced perspective on cloud cover dynamics, with the results shedding light on the evolution of cloud cover over the video's duration.The report meticulously outlines the methodology employed, encompassing data preprocessing techniques, feature extraction, and model training parameters. It presents the findings of the classification process, including accuracy and performance metrics, and discusses insights gained from the analysis of results. This paper contributes to the scientific discourse by presenting a comprehensive framework for cloud analysis in video sequences, with practical implications for a range of disciplines. © 2024 IEEE.