Comparison of ESRI's Global Land Cover Product at Regional Level using Sentinel Imageries and different Machine Learning Algorithms; A Case Study of Nepal

Keywords: Support Vector Machine, Random Forest, Smile Cart, Land Use Land Cover, Accuracy Assessment
Abstract Type: Poster Abstract

Abstract

Accurate land use land cover (LULC) classification plays a vital role in the sustainable management of natural resources and helps to learn how the landscape is changing. The ESA’s Sentinel satellites have laid the foundation for LC products on a global scale. This study aims to evaluate the land cover classification performance of global datasets at the regional level through a case study for three districts (Kailali, Kavrepalanchok, and Jumla) of Nepal for the year 2021 using three supervised classifiers, support vector machine (SVM), classification and regression tree (CART), random forest/tree (RF/RT). The accuracy assessment of these districts revealed that SVM showed an overall accuracy of 91.35%, higher than CART and RF/RT. The user’s accuracy for the image ranged from 57.1% to 97.1%, while the producer’s accuracy varied from 26.7% to 93.6%.However, the Kappa coefficient of 0.88 indicates that the two classifications are in high agreement, meaning that the classification of the image is very accurate. From this, it can be well concluded that Support Vector Machine algorithm gave the best result among all the algorithms used.
Again while comparing the SVM results with ESRI Global LC product to evaluate the performance of the global product at the regional level, the result showed water as most accurately classified (71%) among all the classes whereas barren land had the lowest accuracy of about 6%. This might be because of the reason that our classification used primary data source whereas ESRI used secondary source for training sample data.