A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Remote Sensing and Geographic Information System, School of Engineering and Technology

Thesis (M. Sc.) - Asian Institute of Technology, 2019

This study is aimed to develop an alternative approach of classifying subclass types of urban land-use since the detailed detection of subclass types of urban land-use namely day markets, hotels, night markets, shopping malls, weekend markets, government offices, hospitals, universities, condominiums and villages remains a challenge in remote sensing method. With the proliferation of smartphone devices, social media has been part of human activity. As social media becomes popular, it gives vast amount of information every day from people who shared their sentiments, opinions, new information via different social media platforms. One of the prevalent social media platforms in Bangkok, Thailand is the twitter. Understanding the tweeting frequency of unique users in different time and place gives an essential knowledge that can be used in the classification of subclass types of urban land- use. This study used 4,605,766 tweets with geo-Iocations with 166,511 unique users for twenty three months period which were grouped into different grids of 30m by 30m within Bangkok Metropolitan Area (BMA). The spatio-temporal analysis was conducted and 500 grids were selected to serve as the training and testing datasets with split of 80% and 20%, respectively. A 5-fold stratified cross validation was applied to ensure that each subclass types of urban land-use are equally represented across each test fold. The SVM classifier was used in this study because of its ability to classify non-linear datasets and able to cater high-dimensional data. Also, different kernel functions in SVM were explored and evaluated namely polynomial kernel, Pearson VII function-based Universal Kernel (PUK) and Radial Basis Function kernel (RBF). Based on the result, PUK yielded highest overall accuracy and kappa coefficient with value of 90% and 88.89%, respectively compare to polynomial and RBF kernels. Using this kernel type, the subclass types of urban land-use namely night market, weekend market, shopping mall, government office and university got the highest precision and recall with values above 92% while the other five subclass types namely day market, hotel, hospital, condominium, and village were having low value below 85%. The result shows that twitter can be a potential data source in detailed classification of urban land-use into different subclass types of urban land-use.