Neural Network Based Recognition of Smoke Signatures from Lidar Signals

摘要

The automatic recognition of smoke signatures in lidar signals collected during very small-scale experimental forest fires using neural-network algorithms was studied. An algorithm for pre-processing of raw lidar signals is proposed, which selects suspicious backscattering peaks and makes them unbiased and scale independent. The resulting patterns can be successfully classified as corresponding to alarm or no-alarm conditions by a neural-network algorithm based on a simple one-neuron structure (perceptron). In the case of an alarm, the pre-processing algorithm provides the location of the smoke plume. Five algorithms selected from the literature, and one that was specially developed, all with learning rate adaptation, were used for training the perceptron. Their efficiencies and statistical properties were compared. The best perceptron classifier presented an efficiency of 97% in the classification of smoke-signature patterns and a false alarm rate of 0.9%.