Extreme Learning Machine for classification of high resolution remote sensing images and its comparison with traditional Artificial Neural Networks (ANN)

Abstract

Artificial Neural Networks (ANN) is an important technique for land cover classification of high resolution images. However, there are many inherent limitations of ANN based on Multi-Layer Perception (MLP) with back propagation such as the necessity of fine-tuning the number of input parameters and slow convergence time. Therefore, an attempt has been made to introduce and explore the potential of an Extreme Learning Machine (ELM) which deviates from iterative weight adjustment of neurons during the learning process and is extremely fast at classifying high resolution images. A detailed comparison of the ELM is made with back propagation ANN with better learning algorithms (Scaled Conjugate Gradient (SCG) and Levenberg-Marquardt (LM)) in terms of accuracy, the time required for fine tuning different parameters, and generalisation capability. A high resolution QuickBird satellite image collected over an area of Warsaw, Poland was used for the analysis. Experimental results showed that the ELM produced classification accuracy comparable to that achieved with newer state-of-the-art ANN. The benefits of employing the ELM over conventional ANN are the need of determining only one user parameter, namely the number of neurons in the network, and the significantly lower computational costs. The simplicity of needing to determine only one parameter and the extremely high speed of the ELM could be extremely helpful for different applications when fast but accurate classification is desired.