Deep Learning for Humanitarian Landmine Detection in Aerial Remote Sensing Imagery Using Vision Transformers and Transfer Learning

Landmines remain a significant global threat, affecting human safety and economic stability. Traditional landmine detection methods face challenges, particularly in detecting deeply buried mines. Recent advances in deep learning, particularly Vision Transformers (ViTs), offer promising solutions for automated landmine classification. This research implemented a pre-trained ViT model for landmine detection using aerial infrared images. The classification was conducted in two phases: free zones vs. landmine areas and surface landmines vs. deep landmines. Transfer learning was applied to adapt the ViT model to the dataset, and data augmentation techniques were used to enhance model robustness. The model was evaluated using accuracy, F1-score, sensitivity, specificity, and AUC. Phase one (free zone vs. landmine) achieved 88% accuracy, 87% F1-score, and an AUC of 0.86, with sensitivity and specificity of 33% and 96% for free zones and 96% and 33% for landmines. Phase two (surface vs. deep landmine) achieved 80% accuracy, 80% F1-score, and an AUC of 0.91, with sensitivity and specificity of 82% and 79% for surface and 79% and 82% for deep landmines. This research highlights the potential of ViTs for landmine detection, offering an efficient, automated, and scalable approach. The findings suggested that the deep ViT architecture can enhance demining efforts, reducing human risk and improving detection accuracy. Future work should explore larger datasets, real-world deployment scenarios, and model improvements for better deep landmine detection.