Unsupervised band removal leading to improved classification accuracy of hyperspectral images

Remotely-sensed images of the earth’s surface are used across a wide range of industries and applications including agriculture, mining, defence, geography and geology, to name but a few. Hyperspectral sensors produce these images by providing reflectance data from the earth’s surface over a broad range of wavelengths or bands. Some of the bands suffer from a low signal-tonoise ratio (SNR) and do not contribute to the subsequent classification of pixels within the hyperspectral image. Users of hyperspectral images typically become familiar with individual images or sensors and often manually omit these bands before classification. We propose a process that automatically determines the spectral bands that may not contribute to classification and removes these bands from the image. Removal of these bands improves the classification performance of a well-researched hyperspectral test image by over 10% whilst reducing the size of the image from a data storage perspective by almost 30%. The process does not rely on prior knowledge of the sensor, the image or the phenomenology causing the SNR problem. In future work, we aim to develop compression algorithms that incorporate this process to achieve satisfactory compression ratios whilst maintaining acceptable classification accuracies.