IMPROVING MARKOV RANDOM FIELD BASED SUPER RESOLUTION MAPPING THROUGH FUZZY PARAMETER INTEGRATION
Keywords: Vague Urban Land-cover, SAM, MRF, SRM, Fuzzy mean and Fuzzy covariance
Abstract. The objective of this study was to improve the Markov Random Field (MRF) based Super Resolution Mapping (SRM) technique to account for the vague land-cover interpretations (class mixture and the intermediate conditions) in an urban area. The algorithm has been improved to integrate the fuzzy mean and fuzzy covariance measurements, to a MRF based SRM scheme to optimize the classification results. The technique was tested on a WORLDVIEW-2 data set, acquired over a highway construction area, in Colombo, Sri Lanka. Based on the visual interpretation of the image, three major land-cover types of this area were identified for the study; those were vegetation, soil and exposed grass and impervious surface with low medium and high albedo. The membership values for each pixel were determined from training samples through Spectral Angle Mapper (SAM) technique. The compulsory fuzzy mean and the covariance measurements were derived using these membership grades, and subsequently was applied in MRF based SRM technique. The primary reference data was generated using Maximum Likelihood Classification (MLC) performed on the same data which was resampled to 1m resolution. The scale factor was set to be (S) = 2, to generate SRM of 1m resolution. The smoothening parameter (λ) which balances the prior and likelihood energy terms were tested in the range from 0.3 to 0.9. SRM were generated using fuzzy MRF and the conventional MRF models respectively. Results suggest that the fuzzy integrated model has improved the results with an overall accuracy of 85.60% and kappa value of 0.78 between the optimal results and the reference data, while in the conventional case it was 77.81% of overall accuracy with kappa being 0.65. Among the two MRF models, fuzzy parameter integrated model shows the highest agreement with class fractions from the reference image with a smallest average _MAE (MAE, Mean Absolute Error) of 0.03.