Assessing cloud contamination effects on K-means unsupervised classifications of Landsat data

Satellite data, such as obtained by Landsat 5 or 7 sensors, can be effectively used for large-area land cover classifications. Given that approximately 50% of the Earth is covered in cloud at any time, one of the significant challenges in creating repeatable and robust classifications is to understand and appropriately address cloud contamination in Landsat images. The scope of many of the large area mapping projects and the associated large volumes of data to be processed suggest that unsupervised classifications and automated processes may be necessary to obtain timely results. An experiment was developed to investigate the effect of cloud contamination on unsupervised classifications. It was determined that when a small number of classes are used cloud effects in the cloud-free portion of the scene can often be managed by allocating the majority of clusters to clouds. When a large number of classes are required, clouds significantly skew the non-cloud cluster characteristics.