Buried and surface mine detection from thermal image time series

Kaya, Serkan
Since many countries suffer from existence of landmines in their territory, there is a growing demand for reliable landmine detection systems. Most of these systems require an operator to enter into the minefield. However, infrared sensor methods can be used remotely; hence, they do not put lives at risk during the search operation. In this thesis, a new approach to the infrared sensor method, which gives promising results, is presented.Buried landmines generate specific spatiotemporal thermal image patterns on the surface. Noise-reduced thermal image time series were used after preprocessing with a circularly symmetric filter different from other studies. Supervised classification methods (Support Vector Machine, Mahalanobis Discriminant Analysis, Quadratic Discriminant Analysis, and K-Nearest Neighbor) are applied on the filtered image series.Proposed method gives promising solutions that were verified with enlarged data sets. Different parameters (humidity, burial depths, training sample sizes, time intervals, seasons, spatial filter's sizes etc.) influences on the solutions were examined. To find thevimost useful time intervals, a 2-D heat transfer simulation was performed. Particularly using image series within nighttime, sunrise, and sunset times allow finding four-centimeter deep plastic and 6.8-centimeter deep metal practice landmines with the proposed method. The best solutions were got with the 50-pixel outer diameter spatial filter and the 24-sample on this diameter. Detection rates were calculated higher in quasi-humid soil than dry soil.