A Comparative Assessment of Unsupervised and Supervised Methodologies for LANDSAT 8 Satellite Image Classification

Developing accurate land cover maps is a fundamental prerequisite for natural resource management, environmental modelling and urban planning studies. Several unsupervised and supervised algorithms are available in the literature for classifying LANDSAT satellite images, and selecting an optimum approach is of critical interest. This paper compares two unsupervised (ISODATA; K Means), and three supervised (Spectral Angle Mapping (SAM), Minimum Distance (MD) and Maximum Likelihood) algorithms for classifying a mid-resolution (30 × 30) m LANDSAT 8 satellite image to develop Land Cover (LC) maps for Nashik city in western India region. Post classification stage, sieve filtering and manual corrections are applied for image enhancements. The Kappa accuracy metric is adopted for comparing the accuracy of LC maps against 100 reference ground points using the Google Earth Engine. The Maximum Likelihood algorithm delivered the highest classification accuracy (73.8%), followed by SAM (70.7%), MD (68.1), K means (41.5%) and ISODATA (31.2%) algorithms. Further, accuracy enhancements are attained by the classification sieve filter (83.2%) and by applying manual corrections (89.7%).