Transfer-ensemble based deep learning approach for improved weather image classification task

Abstract

Weather prediction is crucial for various domains, including transportation systems, particularly for autonomous vehicles. Accurate forecasts aid in anticipating traffic flow, ensuring safety, and mitigating the impacts of severe weather on vehicle operations. Early warning systems rely on precise weather forecasts to minimize accidents and congestion, highlighting the importance of predictive technologies in enhancing transportation resilience. In this research work, the goal is to propose a weather phenomenon classification approach using the ensemble techniques and state of the art deep neural networks namely Inception v3, RestNet 50 and VGG16 deep neural network models. The goal is to categorize weather events into 11 distinct classes, leveraging deep learning for improved accuracy. Focusing on outperforming existing models on the Weather Phenomena Dataset (WEAPD), the study employs data augmentation ensemble techniques and meticulous hyper- parameter tuning. The optimized VGG16 model with data augmentation individually achieves an impressive overall accuracy of 92.10%, surpassing other established models. Moreover, a novel ensemble technique of geometric harmonic mean (GHE) of the three deep neural networks (VGG16, RestNet50 and Inception V3) has achieved an overall best accuracy of 92.82% Another approach for making better decision in ensembled models is by using cubic custom mean which gives an average classification of 92.86%. This research contributes to advancing weather event classification accuracy, with potential implications for meteorology and related applications. We have also validated and refined the model through iterative experimentation, ensuring an effective balance between model complexity and efficient training.