Multi-channel descriptors and ensemble of Extreme Learning Machines for classification of remote sensing images

Abstract

We present a method for real-time scene classification which achieves high accuracy without a time consuming descriptor learning step and kernelized classifiers. Robustness of the classification is achieved by combining the powerful multi-channel Gabor-based descriptors and an ensemble of Extreme Learning Machines (ELM). We first extend the recently introduced Binary Gabor Patterns (BGP) to multi-channel images. This is done by extracting BGP over several color channels and embedding an additional compact color layout descriptor. Then we propose an effective method for the aggregation of multiple ELMs into a single classification system, which leads to significant classification accuracy improvements. The experimental evaluation demonstrates that multi-channel color information constantly improves classification results. The integration of multiple ELMs into an ensemble using the proposed aggregation strategy significantly outperforms linear SVM in terms of accuracy, and reaches results similar to the non-linear SVM while operating in real time. Therefore, an ensemble of ELMs with the proposed aggregation strategy could be used as an efficient alternative to the non-linear SVM for remote sensing image classification tasks.