Burned area mapping in Mediterranean environment using medium-resolution multi-spectral data and a neuro-fuzzy classifier

Abstract

In this study, we assess the performance of a self-organising neuro-fuzzy classifier for burned area mapping using multi-spectral satellite data. The proposed neuro-fuzzy model incorporates a multi-layered structure consisting of two types of nodes. The first type is a generic fuzzy neuron classifier (FNCs), whereas the second is solely a decision fusion operator. The Group Method of Data Handling algorithm is used for structure learning providing the model with self-organising attributes and feature selection capabilities. The resulting novel structure consists not only of layers of FNCs but also of layers with only decision fusion due to the nature of the burned area mapping problem. The algorithm is applied to an entire LANDSAT-5 TM multi-spectral image, acquired over central Greece shortly after the major wildfire events of the summer of 2007. In addition to the self-organising neuro-fuzzy classifier, the image data set was classified using neural networks, support vector machines and AdaBoost algorithms. In general, the neuro-fuzzy burned area map presented the highest overall accuracy (more than 95%) compared to the other methods. However, the differences were not statistically significant as suggested by the results of the McNemar's test.

Keywords:

• neuro-fuzzy classification
• decision fusion
• fire scar mapping
• feature selection
• NN
• SVM

Acknowledgements

The Landsat Data was available from the U.S. Geological Survey.