Thermal-Image Landmine Detection using MobileNetV3 and YOLOv8 Ensemble Deep Learning

Abstract

Post-conflict landmine detection requires precise and implementable solutions in resourceconstrained scenarios. In this work, thermal pictures captured by drones with infrared sensors are used to identify landmines using ensemble deep learning. To improve safety-critical reliability, the suggested system combines a lightweight MobileNetV3 classifier with a YOLOv8 object detector. Thermal pictures are preprocessed and enhanced to improve resistance against background, lighting, and mine appearance, and the dataset is split into training, validation, and test sets for fair evaluation. While YOLOv8 is taught to identify suspected mines using bounding boxes, MobileNetV3 is designed for binary mine/no-mine classification by transfer learning. To reduce false negatives in humanitarian demining, an ensemble technique integrates the predictions of both models. Experimental results show that the ensemble outperforms the individual models in terms of safetyoriented parameters for thermal imaging detection quality, accuracy, precision, and recall. The study found that a lightweight ensemble of MobileNetV3 and YOLOv8 on thermal images can improve aircraft landmine identification in real-world circumstances.