Enhanced Landmine Detection in Thermal Images Using MobileNetV3–YOLOv8 Ensemble Learning

Abstract

Landmines continue to pose a substantial concern in post-conflict areas, endangering human life, infrastructure and socio-economic development. Accurate and timely detection of buried or partially exposed landmines is crucial for safe and effective humanitarian demining operations. Conventional detection methods can be labor intensive and need expensive equipment, making them slow, costly and difficult to deploy at scale, or in resource-limited settings. Recent advances in thermal imaging and deep learning have offered potential alternatives for automated landmine identification. In this paper, an ensemble deep learning framework for thermal-image-based landmine detection is proposed with MobileNetV3 and YOLOv8 . Thermal images collected by infrared sensors installed on a drone are preprocessed and upgraded to improve the visibility of the targets and minimize the impacts of environmental fluctuations, background clutter and illumination inconsistencies. The suggested solution uses MobileNetV3 as a lightweight classification model to differentiate between mine and non-mine regions, while YOLOv8 is used for precise object location and detection by predicting bounding boxes. To enhance the detection reliability and limit false negatives, which is a vital need in humanitarian demining applications, an ensemble technique is designed to fuse the merits of both models. Experimental results on thermal-image datasets show that the proposed MobileNetV3–YOLOv8 ensemble outperforms individual models in terms of accuracy, precision, recall, and overall detection performance. The framework shows high durability under the tough real-world situations and effective computing performance for implementation on resourceconstrained platforms. The results demonstrate that the suggested ensemble approach offers a reliable and practical solution for drone-assisted landmine identification and can contribute to safer and more successful humanitarian demining operations.