Cloud Mask Application on Landsat 8 OLI Imagery for Performing Automatic Classification of Vegetation and Water Bodies

In remote sensing specific indices based on appropriate multispectral band algebra formulas allow to highlight differences among land cover types such as vegetation, water bodies and bare soil. The application of an unsupervised classification algorithm on the resulting layer categorizes all pixels in the image, obtaining a given set of labels or land cover themes. Accurate results are achieved even when unsupervised classification is applied to a subset of the available bands, usually the least correlated ones. However, the presence of clouds in the considered scene can affect the classification process and produce unsatisfactory results. This paper highlights how the preventive application of masks on the clouds, avoiding that the pixels belonging to them are subject to classification and contribute to the statistical definition of the classes, eliminates a source of ambiguity for the unsupervised classification. In particular, the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) are applied to Landsat 8 OLI multispectral images. The resulting layers, as well as two subsets of available bands, i.e. Coastal/Aerosol + Near-infrared + Short Wavelength Infrared and Blue + Green + Red + Near-infrared bands, are submitted to the ISODATA unsupervised classification algorithm. The use of confusion matrices and related evaluation indices shows that, for the two considered band subsets, the thematic accuracy of the resulting land use map is higher when cloud masks are introduced, as testified by an Overall accuracy greater than 0.98. On the other side, NDVI and NDWI submitted to ISODATA provide high performance both in the presence and absence of cloud masks: the results confirm the usefulness of the band ratio that tends to level the disturbance actions, in this case, also that due to the presence of clouds.