2.3 Image classification and accuracy assessment
A modified version of the Anderson level 1 land use and land cover (hereinafter, LULC) classification scheme (Anderson et al., 1976) is used to classify Sentinel data into discrete LULC categories (Table 2). A hybrid approach, comprising unsupervised and supervised techniques, is employed (Bauer et al., 1994). First of all, an Iterative Self Organizing Data Analysis (ISODATA) algorithm is used to derive signatures from multitemporal Sentinel data, pertaining to the study area. The signatures were then evaluated using histogram and transform divergence (TD) techniques to ensure normality (Yuan et al., 2005). The TD value of ≥ 1900 is accepted in this work. Besides, reference data (e.g., RapidEye, World View 2 and Google Earth Images) for each year is considered, side by side, in an image processing system to determine the usefulness of individual signatures. This process helps isolating signatures that are suitable for classifying images. A maximum likelihood routine is subsequently applied to derive distinct LULC categories (Bolstad & Lillesand, 1991). Since the study area has diverse land covers, misclassification of pixels is noticed between shrubs and agriculture, mixed forest, and canopy trees and homestead vegetation cover. To subdue issues with misclassification, post classification refinement is carried out to recode mixed pixels into correct LULC categories (Harris & Ventura, 1995). Finally, three maps of the study area are obtained, representing LULC data of 2017, 2018 and 2019.
To evaluate classification accuracy, 100 points for each LULC classes are derived from high resolution images, noted above (Table 1), with a stratified random sampling technique. Using reference data and classified images, an error matrix is then prepared from which four accuracy metrices (e.g., overall, producers, user’s accuracies and kappa statistics) are computed.