Weather image classification using EfficientNet and Dual Attention Block

Classification of Weather images is the dominant research task in weather classification. Especially for applications like climate forecasting, intelligent traffic signaling, and autonomous driving systems, utilizes weather type prediction as the primary task. Weather image classification is an important task due to the diverse properties of weather conditions. We propose an EfficientNet and Dual Attention Block model to perform five-class weather image classification. We use a pretrained EfficientNet to extract features from weather images. The features are re-convoluted using a dual attention block, and a global pooling followed by a dense network for the final classification. The proposed model is evaluated using a publicly accessible weather image dataset from kaggle consisting of 1530 images. Our proposed model outperforms existing pre-trained models with 96.73% of accuracy.