학술논문
Advancements In Passive Landmine Detection A Multiclass Approach With Fluxgate Sensor And Machine Learning Models
Document Type
Conference
Author
Source
2024 41st National Radio Science Conference (NRSC) National Radio Science Conference (NRSC), 2024 41st. 1:158-165 Apr, 2024
Subject
Language
ISSN
2837-018X
Abstract
This paper addresses the pressing issue of passive landmine detection by merging the capabilities of magnetometers and sophisticated machine learning methods. The accurate detection and categorization of landmines are crucial to mitigate the severe consequences of these devices in areas of conflict and post-conflict. Passive detection techniques offer significant benefits over active methods, they reduces the risk of detonation during detection, enhancing safety for operators and the surrounding environment. This paper presents an in-depth analysis of the methodology for passive detection of landmines by fluxgate magnetometers. A detailed literature review is conducted to investigate current techniques, highlighting both the progress and constraints of passive detection strategies. Importantly, the study emphasizes the essential role of machine learning in enhancing passive landmine detection. Three unique machine learning models—Classification and Regression Tree (CART), Support Vector Machine with Error Correcting Output Codes (ECOC-SVM), and Artificial Neural Network (ANN)—are utilized to process and model data from landmine fluxgate magnetometer detection. By leveraging Bayesian optimization and cross-validation, the research develops efficient and generalizable models, significantly improving the precision of passive landmine detection and classification. Bayesian optimization was applied using six different acquisition functions including expected improvement, lower confidence bound, and probability of improvement. The performance of the three models was assessed and contrasted using the general accuracy of the confusion matrix for both training and test subsets. The Artificial Neural Network (ANN) model demonstrated superior performance, achieving an accuracy of 100% on the training set and 82% on the test set. The study also includes a visualization and analysis of the ANN model’s partial dependence on the voltage of the fluxgate sensor.