Image-Driven Temperature Prediction with Optimized ResNet50 Model

This study proposes a new deep learning-based approach for mean temperature prediction based on weather attribute images using an optimized ResNet50 convolutional neural network model. A dataset of curated diverse weather type images, such as snow, rain, and fog, annotated with corresponding mean temperature labels, was used to train and evaluate the model. The methodology incorporates advanced techniques such as data augmentation and transfer learning to enhance the prediction accuracy and robustness of the model. ResNet50 was further fine-tuned with additional layers designed for regression tasks to enable it to correctly relate image feature space and temperature values. Rotation, zoom, and varying brightness are some of the data augmentation techniques applied as countermeasures against changes in lighting, time of day, and orientation in images. The model achieved a Mean Absolute Error (MAE) of 2.54°C and a Root Mean Squared Error (RMSE) of 3.12°C during testing, demonstrating competitive predictive capability. This research has the potential to combine deep learning with weather imagery for temperature prediction and is promising for areas like remote sensing, mobile climate applications, and disaster management systems where traditional sensor data may be limited or unavailable.