EXPLORING EXTREME LEARNING MACHINES FOR IDENTIFYING ROADS IN LANDSAT-4 THEMATIC MAPPER IMAGERY

Abstract

Identifying roads in remotely sensed imagery provides one indicator land usage for agricultural and urban planning. The data collected from one early sensing platform, the Landsat-4 Thematic Mapper, continues to be useful as historic data for land change assessment. This work describes the result of applying the Extreme Learning Machine algorithm to train a Single-Layer Feed-Forward Neural Network to detect roads in rural imagery. The results are then compared with a Backpropagation-trained network, Learning Vector Quantization, and a Pulse-Coupled Neural Network. Results indicate that the Extreme Learning Machine is competitive with other approaches, and offers the advantage of rapid training.

Index Terms - Extreme Learning Machine, Neural Networks, Landsat, Remote Sensing, Computing