Modified possibilistic c- means with constraints (MPCM-S) approach for incorporating the local information in a remote sensing image classification.

Abstract

In this research letter a Modified Possibilistic c-Means with constraints (MPCM-S) algorithm has been proposed for an image classification. This algorithm has been presented to overcome the shortcoming of Possibilistic c-Means (PCM) and Modified Possibilistic c-Means (MPCM). It is done by incorporating the local information using the spatial constraints and expected to control the effect of neighbouring terms. This Neighbourhood labelling has been introduced in MPCM-S by introducing local window (NR) including noise minimizer (λi) and regularizer parameter (ηi). Experiments have been conducted on the Formosat-2 and Landsat-8 satellite imagery of study area Haridwar, India. The classified outputs generated for both types of satellite data by applying the algorithms viz., PCM, Possibilistic c-Means with constraints (PCM-S), MPCM and Modified Possibilistic c-Means with constraints (MPCM-S) were optimized using an independent assessment method known as Mean Membership Difference (MMD). Root Mean Square Error (RMSE) has been conducted to test the proposed algorithms. RMSE has been computed between original and noisy classified image for Formosat-2 satellite imagery in which MPCM-S produces least RMSE. Furthermore, the Fuzzy Error Matrix (FERM) based assessment of accuracy method has been used to evaluate the classified outputs from Landsat-8 data using the Formosat-2 as reference data. The proposed algorithm has also been able to extract heterogeneously distributed land cover classes. The results of this study show that proposed MPCM-S classifier is effective in minimizing noisy pixels and outliers and maximize the fuzzy overall accuracy than PCM and MPCM.