Remote sensing image classification for forestry using MRF models and VQ method

Conventional pixelwise classification techniques have two drawbacks. The first is that isolated misclassifications tend to occur because they classify each site independently using spectral information only. The second is that sample data which represents each class are indispensable to estimate model parameters or to train classification neural networks. In this manuscript a new unsupervised contextual method is proposed for multispectral remote sensing image classification. Markov random field (MRF) models are used for modeling the observed and classified images in the proposed method. Both spectral and spatial information can be exploited by the MRF models so as to dissolve contextual inconsistency caused by the first drawback. In addition the vector quantization (VQ) method is introduced to dissolve the second drawback. The VQ method classifies the observed data into several clusters without using any sample data. The image classified by the VQ method is so coarse that it includes misclassification and unknown-class sites, however it can be used as an initial classified image for the following MRF-based classification algorithm. The proposed method was applied to LANDSAT Thematic Mapper data to discriminate deciduous trees from ever-green trees. The accuracy of classification was 64.4% by the VQ method and it was improved up to 88.8% by the MRF-based method.