Study of Statistical and Computational Intelligence Methods of Detecting Temporal Signature of Forest Fire Heat Plume from Single-Band Ground-Based Infrared Video

This thesis will analyze video from land-based, cooled mid-wave infrared cameras to identify temporal features indicative of a heat plume from a forest fire. Desirable features and methods will show an ability to distinguish between heat plume movement and other movements, such as foliage, vehicles, humans, and birds in flight. Features will be constructed primarily using combinations of statistics and principal component analysis (PCA) with intent to detect key characteristics of fire and heat plume: persistence and growth. Several classification systems will combine and filter the features in an attempt to classify pixels as either heat or non-heat. The classification systems will be tuned and compared with common metrics of error rate and computation time. It was found that the movement pattern of a heat plume could be distinguished from the similar movement pattern of foliage by detecting outlier movement patterns, a phenomenon associated with the growth property of fire. Outlier movement patterns were best detected by thresholding the quotient of mean and median of a set of variance measurements over time. The best tested classifier in terms of minimizing false positives without losing the heat signal came from PCA of a dual-range moving average difference.