Texture-based classification of cloud and ice-cap surface features

Abstract

It is well known that the interpretation of high resolution (<100 m) visible and near infrared (e.g. Landsat) imagery of large ice masses is hindered by the uniform reflectivity of snow, ice and cloud surfaces. Such interpretation is at present largely performed manually, but there is a good prospect that it could be automated by the incorporation of image texture. This paper describes preliminary work towards the identification of the most appropriate texture technique, or combination of techniques, and assesses the likely performance of such methods. Different textures are identified with different types of surface cover, and the use of these differences to classify images is investigated. Specifically, we compare a traditional texture measure, the Grey Level Co-occurrence Matrix (GLCM), with a modification of a relatively new technique, fractional Brownian motion (FBM). These two methods are applied to three Landsat MSS images of the Nordaustlandet ice cap, Svalbard. The classification accuracy, computation time and memory required, advantages and limitations of the two methods are compared. The GLCM technique appears to be able to distinguish three groups of image classes, namely dry snow, wet snow, and melt features, ablation areas or cloud cover. The FBM technique is computationally more efficient, and though it performs in general less well than the GLCM technique it gives better discrimination of cloud cover.