Abstract
Weather conditions have a significant effect on humans' daily lives and production, ranging from clothing choices to travel, outdoor sports, and solar energy systems. Recent advances in computer vision based on deep learning methods have shown notable progress in both scene awareness and image processing problems. These results have highlighted network depth as a critical factor, as deeper networks achieve better outcomes. This paper proposes a deep learning model based on DenseNet-121 to effectively recognize weather conditions from images. DenseNet performs significantly better than previous models; it also uses less processing power and memory to further increase its efficiency. Since this field currently lacks adequate labeled images for training in weather image recognition, transfer learning and data augmentation techniques were applied. Using the ImageNet dataset, these techniques fine-tuned pre-trained models to speed up training and achieve better end results. Because DenseNet-121 requires sufficient data and is architecturally complex, the expansion of data via geometric augmentation-such as rotation, translation, flipping, and scaling-was critical in decreasing overfitting and increasing the effectiveness of fine-tuning. These experiments were conducted on the RFS dataset, and the results demonstrate both the efficiency and advantages of the proposed method, which achieved an accuracy rate of 95.9%.