Weather recognition based on still images using deep learning neural networks

Abstract

Weather recognition from still images remains a challenge due to weather diversity and lack of distinct characteristics amongst weather conditions. The building blocks of deep learning involves a lot of hyperparameters, which is difficult to tune manually. Fine-tuning Hyperparameter is a crucial part of building a good deep learning model. This study proposes a simplified ResNet-15 model fine-tuned by two distinct optimisers, SGD, and Adam for the purpose of optimising the momentum, number of dense layers, learning rate, batch size and dropout rate to find the optimal hyperparameters that gives the best performance on the model. The Convolutional layers are used to extract the most related visual features, then the images are classified through the fully connected layers of the Softmax classifier. To evaluate the recommended approach, a comparison of hyperparameter tuning with and without hyperparameter tuning of ResNet-15 during the experiments shows that fine-tuning of ResNet-15 hyperparameter using random search optimisation method gives more accurate result with accuracy of 97.29% than other techniques.