Today’s technology allows for classification of ground cover using remote sensing imagery. While simple color or black and white images were originally used, advances in sensor technology now allow for remote sensed data to be collected as tens or hundreds of images corresponding to narrow adjacent light wavelength intervals. Such data, traditionally called multispectral or hyperspectral, reveals minute differences between materials, and leads to increased classification precision.
Often some of the spectral images collected have high signal-to-noise ratio that, in fact, reduces accuracy. This paper suggests an approach to deal with such bands, by identifying them and eliminating them prior to the classification process. The proposed approach clusters the data, and then computes the distances between the clusters on each band using a modification of a traditional distance (Bhattacharyya). When the maximum cluster distance is below a threshold value, the band is eliminated. The method not only reduces the data set, but also, according to the experiments, leads to improved classification.
The paper is well written and easy to read. Given the venue where it was presented, the Australasian Computer Science Conference, it may not have received the visibility it deserves from the remote sensing community. This would be a loss, since the technique presented is simple and inciting. Unsupervised band reduction is an important problem in the field, and the results suggest the method performs well. Unfortunately, the data set used is not very appropriate for in-depth testing. While a significant study has been conducted, the ground truth data needed for validating the classification lacks accuracy. Nevertheless, the approach warrants consideration, and I hope that future investigation on other data sets will provide a better understanding of its efficiency.