MACHINE LEARNING BASED TECHNIQUES TO EXTRACT URBAN AREA USING SENTINEL-2A DATA

Abstract

Urbanization is the major cause of concern nowadays for the whole world as it is increasing at a very tremendous rate and thus, has to be considered for studies to deal it at its best. Several studies have already been conducted and new researches still going on in this particular field. Excess urbanization resulted into several negative impacts like land subsidence, land deformation, unplanned city development, lack of resources to sustain the life, etc. To consider optical data for built-up area mapping is not so easy using conventional supervised classification methods. But, here in the thesis work, several machine learning algorithms have been applied so as to demarcate the boundary between built-up and other classes. In this way, these algorithms work effectively to extract the built-up class from the study area. For the study, multispectral band dataset of Sentinel-2A satellite is considered so as to get the accurate extraction of the urban area around the Roorkee, Haridwar in India. The bands useful to get the information of built-up area have been used which include bands 2, 3, 4, 5, 6, 7, 8, 8A, 11 and 12. The Sentinel-2A multispectral data has been pre-processed on QGIS software for atmospheric correction and reflectance is obtained for further work. Using the pre-processed dataset, the whole work has been conducted on R-Programming software. On the basis of mean spectral reflectance values for each of the class/labels considered on the satellite image, training dataset is created manually as a shape file (.shp). The spectral profile for different classes was also studied to differentiate the spectral signatures. On that basis, the classes in training dataset apart from built-up merged into one and labelled as ‘Non- Urban’ using ERDAS software and built-up class is categorized as ‘Urban’. NDVI (Normalized Difference Vegetation Index) spectral index is also applied on the satellite imagery to analyse the different features like water, vegetation and built-up separately and which was further used to verify our results obtained from the machine learning algorithms to some extent. But, due to intermixing of pixels in the scene, images obtained after applying spectral index was also not fully accurate in terms of differentiating the classes. Several machine learning models such as Decision Tree, Random Forest (RF), Support Vector Machine (SVM) and Neural Network (NN) models have been trained using the training dataset and applied over the testing dataset to predict the overall accuracy and kappa value to evaluate the best model for the work. The overall accuracy for the above models is in preferring order as 94.50, 93.00, 92.00 and 91.5 % respectively for SVM, RF, NN and Decision Tree.