Geographic Object-Based Image Analysis for Small Farmlands using Machine Learning Techniques on Multispectral Sentinel-2 Data

Abstract

This study intends to classify the land cover of an area especially small farmlands using object-based image analysis (OBIA) method and evaluates the performance of a supervised classifier. Multi-spectral Sentinel-2 imagery which is freely available is used and four supervised classifiers are applied to it. The study area was divided into four major classes namely Urban, Wheat, Tobacco and Other Vegetation with varying accuracy values. The imagery was first resampled to 10 m spatial resolution and then NDI45 is layer stacked to it. A widely used MRS technique is used for delineating the objects in the imagery. Finally, classification is done through four supervised classifiers k-Nearest Neighbour (kNN), Bayes classifier, Decision Tree (DT), and Random Forest (RF). The accuracy was evaluated through a confusion matrix. The results show that Sentinel-2 imagery is capable of producing thematically detailed land cover maps via the Geographic Object-Based Image Analysis (GEOBIA) approach with accuracies of k-NN 95%, Bayes classifier 88%, DT 81% and RF 79%.