Analysing the Impact of Training Sample Size in Classification of Satellite Imagery

Land Use and Land Cover (LULC) classification is an essential application of remote sensing that allows distinguishing between different land use/ land cover types, like built-up areas, forest lands, wastelands, and agriculture areas. A LULC map can help in better understanding of a region which can help in better decision making. This research utilizes satellite imagery to produce the LULC maps and evaluates several Machine learning algorithms, including Support Vector Machine (SVM), Random Forest (RF), and Classification and Regression Trees (CART), in classifying satellite imagery. This study demonstrates that the performance of these classifiers is strongly interconnected with the training sample size. While the general agreement is that larger sample sizes typically yield better classification, the implications and use of larger datasets can be challenging, particularly because larger sample sizes require more time and resources, especially when ground observations are needed. The study also examines how different machine-learning algorithms perform with various training sizes, focusing on their accuracy and reliability.