A deep transfer learning based convolution neural network framework for air temperature classification using human...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Environmental Science

Background:

• Weather recognition is vital for daily life, impacting prediction, monitoring, and energy.
• Existing image-based weather recognition methods lack accuracy and cover limited phenomena.
• Accurate air temperature classification from visual cues is an underexplored area.

Purpose of the Study:

• To develop and evaluate a transfer learning Convolutional Neural Network (CNN) framework for classifying air temperature levels using human clothing images.
• To assess the performance of various deep transfer learning models, including DeepLabV3 Plus, BigTransfer (BiT), Vision Transformer (ViT), ResNet101, VGG16, VGG19, and DenseNet121.
• To validate the model's interpretability using Gradient-weighted Class Activation Mapping (Grad-CAM).

Main Methods:

• A novel Human Clothing Image Dataset (HCID) with 10,000 images across High and Low air temperature categories was curated.
• A CNN framework leveraging deep transfer learning approaches was implemented for image classification.
• Models were rigorously evaluated using metrics like accuracy, precision, F1-score, recall, AUC-ROC, and confusion matrices.

Main Results:

• DenseNet121 demonstrated superior performance, achieving an accuracy of 98.13% in classifying air temperature levels.
• All evaluated models showed promising results in distinguishing between high and low temperature conditions.
• Grad-CAM analysis provided insights into the visual features driving the classification decisions.

Conclusions:

• The proposed transfer learning CNN framework effectively classifies air temperature levels from human clothing images.
• DenseNet121 offers a highly accurate solution for this task, outperforming other tested models.
• This research presents a valuable tool for enhancing weather prediction and environmental monitoring systems.