Hyperspectral image classification using random forests and neural networks

Abstract

Spectral unmixing of hyperspectral images are based on the knowledge of a set of unknown endmembers. Unique characteristics of hyperspectral dataset enable different processing problems to be resolved using robust mathematical logic such as image classification. Consequently, pixel purity index is used to find endmembers from Washington DC mall hyperspectral image dataset. The generalized reduced gradient algorithm is used to estimate fractional abundances in the hyperspectral image dataset. The WEKA data mining tool is selected to construct random forests and neural networks classifiers from the set of fractional abundances. The performances of these classifiers are experimentally compared for hyperspectral data land cover classification. Results show that random forests give better classification accuracy when compared to neural networks. The study proffers solution to the problem associated with land cover classification by exploring generalized reduced gradient approach with learning classifiers to improve overall classification accuracy. The classification accuracy comparison of classifiers is important for decision maker to consider tradeoffs in accuracy and complexity of methods.