Neural Networks Approach to Fire Severity Mapping from a Single Post-Fire Landsat 7 ETM+ Imagery

Abstract

The objective of this paper was to investigate the potential of a neural network (NN) technique for the delineation of a fire severity map with a single post-fire Landsat 7 ETM+ imagery of the Kang-Won coastal ecoregion of S. Korea. Tasseled Cap (TC), Principal Component (PC), and Intensity-Hue-Saturation (IHS) transforms with MLC (Maximum Likelihood Classification) was used as traditional methods for the comparison. The architecture of NN used has a multi-layer, feedforward type, and employs the modified Levenberg-Marquardt backpropagation algorithm. The NN result outperformed the other methods with a higher classification accuracy of l0%~30%, although it showed only significant difference with the result of IHS in Kappa values. However, this study showed that a neural network technique was better in terms of accuracy and processing efficiency than other techniques when analyzing spatially and spectrally complex patterns resulting from fire on rugged terrains in S. Korea.