Object Based Image Classification for Sentinel 2 Imagery using Machine Learning

(Conference presentation at Canadian Symposium of Remote Sensing 2023)

This project aimed to evaluate the applicability of object-based image classification (OBIA) for mapping land use/land cover (LULC) in dense, heterogeneous urban environments using open-source tools. The study focused on the City of Waterloo, Ontario, an area experiencing rapid urban growth. Sentinel-2 MSI imagery from 2021 (10m resolution) was used alongside 12 cm resolution aerial photography for ground truthing.

A novel classification pipeline was developed entirely using open-source platforms, including QGIS, Orfeo Toolbox (OTB), and Sen2Cor. A multi-index raster (MNDWI-NDTI-NDVIre) was created to enhance class separability, and segmentation was performed using OTB's LargeScaleMeanShift (LSMS) algorithm. The classified LULC maps -comprising five classes: buildings, impervious surfaces, water, vegetation, and bare land - were generated using two machine learning algorithms: Support Vector Machine (SVM) and Random Forest (RF).

Accuracy assessments showed that the SVM classifier achieved an overall accuracy of 87.6% and a kappa coefficient of 0.825, while RF yielded 86.2% accuracy with a kappa of 0.805. The project demonstrated that open-source OBIA pipelines can produce high-accuracy LULC maps suitable for municipal planning and monitoring, while avoiding the salt-and-pepper noise typical of pixel-based methods.