Neural networks for supervised classification of lidar signals at forest-fire surveillance

Abstract

Detection of smoke plumes using lidar (laser radar, light detection and ranging) provides many advantages with respect to passive methods of forest-fire surveillance. However, great sensitivity of the method involves detection of many spurious targets. To be efficient and automated, the lidar-assisted fire detection must be supplemented with effective algorithms of separation of the smoke-plume signatures from the irrelevant peaks due to various atmospheric phenomena and electronic noise as well as from the signals resulted from other targets. This work proposes a new simple and robust algorithm of lidar-signal classification based on the fast extraction of sufficiently pronounced peaks followed by their classification with a perceptron. Within the framework of the radial-basis-function-network ideology, the perceptron capability is enhanced and linear degeneracy overcome by a special fast and highly nonlinear transformation that efficiently increases the number of the perceptron nodes and, consequently, the number of adjustable interconnection weights responsible for the memorization capacity. Application of this method made it possible to develop software for automatic smoke recognition with an error rate as small as 0.31% (19 misdetections and 4 false alarms at recognition of a test set of 7409 peaks).