A Comparative Study of Machine Learning and Deep Learning Models for Weather Image Classification

The study of identifying and categorizing weather images is closely connected to our daily routines and the economy. It affects decisions such as whether to bring an umbrella for the day or the agricultural earnings of an entire country for the year. However, the complexities within the domain of weather classification are manifold, owing to the subtle nature of weatherrelated image features and the multitude of factors contributing to image variability. In this context, we explore the efficacy of three distinct deep learning models: Convolutional Neural Networks (CNNs), Visual Geometry Group (VGG) networks, and Vision Transformer (ViT) models, to address the challenges posed by ambiguous attributes of weather images. The above models exhibit notable differences in their architectural nuances. CNNs excel at capturing local features through convolutional layers. VGG networks are adept at learning intricate feature hierarchies via their deep architecture. ViT, leveraging self-attention mechanisms, excels in capturing global context from images. These models collectively tackle the issue of indistinct weather features by extracting and learning relevant patterns at different scales and levels of abstraction. Our experimentation employs a diverse weather dataset encompassing ten categories, including sunny, cloudy, rainy, snowy, and others, each comprising 500 images. The results are promising, with all three models achieving classification accuracies exceeding 90%. This study underscores the potential of CNNs, VGG networks, and ViT models in effectively categorizing weather images despite the inherent challenges, contributing to improved weather understanding and forecasting, and ultimately benefiting various sectors of society.