Pixel-Based vs Patch-Based Classifiers for the LULC Classification
Keerti Kulkarni, Priyadarshini K Desai · International Journal of Engineering Trends and Technology · 2024
A vast majority of the policy decisions by the civic authorities depend on the geography of the given area. The land use and the land cover of the area also influence the infrastructure facilities availed by the civic bodies. The problem with LULC mapping of the urban areas is the class imbalance. Very few existing algorithms take into consideration this class imbalance. The novelty of this work is the handling of this class imbalance at two levels: the data level and the classifier level. At the data level, uniformly sampled training samples and area-proportional training samples are considered and compared. At the classifier level, pixel-based and patch-based classifiers are considered and compared. A pixel-based Parametric Classifier (Maximum Likelihood Classifier), trained on uniformly sampled training samples, gives an overall accuracy of 73.21% and an overall accuracy of 75.81% when trained on area-proportional training samples. Pixel-based Non-parametric Classifier (multiclass Support Vector Machines), trained on area-proportional training samples, gives an overall accuracy of 83%. The study area is the Bangalore Urban District, and the remotely sensed images are from LANDSAT-8. Patch-based convolutional neural networks give a superior accuracy of 91.2%. Hence, for an imbalanced class dataset, a classifier-level approach (convolutional neural networks) works better, as they look at patches of images rather than individual pixels.