Weather Categorization Using Foreground Subtraction and Deep Transfer Learning

Abstract

We propose a novel foreground subtraction algorithm to extract the Sky pixels and use transfer learning on these processed images using pre-trained ResNet-50 and DenseNet-161 models to classify them. For this study, the Image2Weather’s 2 category weather dataset having 1500 images, is used. By using transfer learning and training, only the last layer by freezing rest of the layers, we have found that the proposed ResNet-50 and DenseNet-161 architectures achieved the classification accuracies of 94.3% and 92.68%, respectively, on the test images and outperformed the state-of-art methods. These results are extremely good as there is a significant improvement of 6–8% in accuracy over the existing models due to foreground subtraction. The proposed method can be implemented in real-time applications to help farmers plan their cropping cycle and for efficient weather forecast activities.