Land Cover Classification of Deception Island with SAR and Optical Images

Optical images are widely used for land cover identification purposes, although they have flaws such as being unusable at night or in the presence of clouds. SAR images, on the other hand, do not depend on solar illumination as they work with microwave pulses, and therefore overcome these problems. They are, however, more difficult to interpret, firstly because they differ from what the human eye is accustomed to seeing, and secondly because they present noise such as speckle which diminishes the quality of the image. There are studies on their use in urban contexts and above all based on multiple polarimetry images, which are easier to process, but there are only few studies based on single polarimetry images and acquired in more complicated contexts, where contrasts and geometries are less marked than in urban contexts. The aim of this work is therefore to classify an optical image, acquired by the Sentinel 2 satellite, and a single polarimetry SAR image, acquired by Sentinel 1, depicting a non-urbanized area, Deception Island, and in temporal proximity: one the day after the other, so as to be able to compare the results, in order to evaluate the possible use of SAR images to replace optical ones when conditions render the latter unusable. The classifications made will be both unsupervised and supervised, pixel-based in the case of the optical image and object-based in the case of the SAR image; the ISO-cluster, Maximum Likelihood, Random Trees and Support Vector Machine classifiers will be used. In the case of the optical image, a classification based on the use of indices will also be added, as this field that has been little studied and could be further explored. The ground truth will be derived from the combination of two sources: the 2005 topographic map produced by the Spanish Army and the Pixel QA Band of a Landsat image close to those of Sentinel 1 and 2 analyzed. The results obtained with the optical image were good, but affected by cloud cover, while the ones of the SAR image were worse due to speckle noise, and the solution of repeatedly applying filters to reduce it actually led to a loss of information that caused the goodness of the classifications to deteriorate. Classification based on the successive use of indices was the one that gave the best results, but not all of the generated classes could be evaluated because of the lack of reference sources (for example, it’s difficult to find a source which states the snow cover each day), so further investigation in this field is encouraged.