Fuzzy c-Means Classification of Multispectral Data Incorporating Spatial Contextual Information by using Markov Random Field

Disclaimer This document describes work undertaken as part of a programme of study at the Indian Institute of Remote Sensing and International Institute for Geo-information Science and Earth Observation. All views and opinions expressed therein remain the sole responsibility of the author, and do not necessarily represent those of the institute. Abstract Remote sensing technologies provide a unique opportunity to map the real world phenomena in a much faster and economic way in comparison to the traditional ground survey methods. The continuous nature of geographical phenomena throws a typical challenge to prepare landuse/landcover maps from remotely sensed data. Very often landcover class changes gradually from one to another, therefore in such condition it is difficult to define sharp boundaries between two landcover classes and fuzzy classification techniques can be used to represent such conditions. However the Fuzzy c-Means classifier (FCM), the most common fuzzy classification technique, does not incorporate the spatial contextual information, which can be useful for further improvement in fuzzy classification results. Markov Random field (MRF) is a mathematical toolbox which characterizes the spatial contextual information in terms of smoothness prior assumption and incorporation of contextual information helped to improve the classification result for hard classifiers. In the present study an algorithm called contextual FCM classifier was developed by using the MRF model and its performance was justified in comparison to the standard FCM algorithm in the context of wetland mapping. The contextual FCM and FCM classifiers were used on AWiFS and LISS-III data with different spatial resolutions i.e. 60m and 20m respectively. For the purpose of validation soft reference data was generated from fine resolution LISS-IV data (5m) using Support Vector Machine (SVM) classifier. The applicability of Euclidean and Mahalanobis norm in contextual FCM classifier were also judged in the context of wetland mapping. The results of different classification techniques were validated using seven different accuracy assessment tools, namely, Root Mean Square Error (RMSE), Pearson's Product Correlation Coefficient (r), Fuzzy Error Matrix (FERM), Sub-Pixel Confusion Uncertainty Matrix (SCM), MIN-PROD, MIN-MIN and MIN-LEAST composite operators with respect to the soft reference data. The results suggest that proposed contextual FCM classifier can improve the fuzzy classification results by incorporating spatial contextual information for remotely sensed data. Therefore it was also found that for contextual FCM classifier Euclidean norm performs better in comparison to the Mahalanobis norm. Several experiments were performed to establish the suitable Simulated Annealing (SA) technique for …