Analysis of support vector machine and maximum likelihood classifiers in land cover classification using Sentinel-2 images

Abstract

Remote sensing data has been widely applied to classify the land cover more frequently and on a near real-time basis for updating as it is more economic, less time consuming compared to ground based survey. Accurate classification of the land use/cover classes such as water body, cropland, built-up area, scrub land, fallow land, forest etc., is one of the biggest challenges in natural resource inventory, management and monitoring. As accuracy of remote sensing data classification is affected by many parameters which include type of data, presence of heterogeneous landscapes in study area, classification approaches etc., as satellite imagery is complex in nature. Many classifiers have been developed and tested on remotely sensed data for better classification. Classification of remote sensing data is mainly divided into two categories such as supervised and unsupervised. In supervised classification, the decision boundaries in feature space are determined by training the samples. Two supervised classifiers namely maximum likelihood (ML) classifier, which is a parametric classifier that assumes data to be normally distributed, and support vector machine classifier (SVM) which is a non-parametric classifier are used in classification. In the present study, the accuracy of these two classifiers is studied on five different data sets of Sentinel-2 satellite image of different years and sessions to accommodate intra and inter annual variations of the datasets. Sentinel-2 satellite images covering part of Nagpur, located in Maharashtra, India were used for the classification. Classifier accuracy has been calculated using overall accuracy and kappa statistics based on ground truth information. The result obtained were carefully examined by comparing classification accuracies and then by visual analysis. The result shows that SVM classifiers gives better overall accuracy and kappa coefficients and its average value for intra and inter annual classification outputs were 91.78% and 0.89 in that order which is far better than ML classifier which gave 87.07% and 0.83 respectively. The experimental results obtained from the present study, it is clear that SVM classifier produced better accuracy than ML classifier in classifying Sentinel-2 optical image and have significant potential in classifying various land cover classes in the heterogeneous land use/land cover conditions of the tropical regime.