Study of Training Parameters Effect in Noise Clustering Classifier for Handling Heterogeneity Within the Class for LULC Classification

A conventional Noise Clustering (NC) algorithm relies on statistical variables such as means or variance–covariance, which are calculated from sets of training sample data. The NC classifier has been experimented in this research work while using a training sample as an ‘individual sample as mean’, to handle heterogeneity within the class. The six categories analyzed included dense forest, eucalyptus, sand, water, wheat, and grassland, utilizing multispectral imagery from LANDSAT-8 and FORMOSAT-2 in the Haridwar region. The classified data utilized the LANDSAT-8 image, whereas the reference data relied on FORMOSAT-2. This study observed that the NC classifier outperformed when applied to the training sample as ‘individual sample as mean’ instead of ‘mean’. Overall Accuracy was calculated through FERM and RMSE methods, and variance was measured to assess heterogeneity within the class from outputs. The overall accuracy of NC applying the training sample as ‘individual sample as mean’ was 86.49%, while using the training sample as ‘mean’ was 82.55%.