Weather identification using models based on deep learning

Abstract

Accurate weather forecasting is increasingly crucial as climate change intensifies the unpredictability of weather patterns, posing challenges to traditional forecasting models reliant on human observation or numerical methods. Researchers are working on precise weather forecasting to improve our preparedness, enabling fast response to any disaster. Among other techniques, deep learning is a prudent method to predict weather forecasts since it can automatically learn and train from a vast amount of data to generate and portray accurate features of an incident. This study evaluates deep learning techniques for weather forecasting based on different meteorological characteristics. This paper examines a few weather variables to evaluate the prediction performance of several deep learning solutions using TensorFlow and pre-trained Keras applications models. For this purpose, the top ten accuracy-based deep learning model architectures have been investigated and evaluated. The operation of each model is distinct. Models like EfficientNetB7, ResNet, MobileNet, VGG19, Xception Inception, ResNetV2, and VGG16 employ a combination of image classification and deep learning models to predict the weather. The WEAPD dataset of 6877 images representing 11 weather phenomena categories was utilized, and the models were trained and validated using an 80:10:10 split. Predictions, extraction of features, and fine-tuning of models were achieved with an accuracy of up to 83.39%. Most models performed well in image classification, enhancing the proposed framework and achieving significant precision in generating weather photos and reports.