Hyperspectral Image Classification Based On Conditional Random Fields Model

Starting from the imaging mechanism of remote sensing, an analysis of the mathematical properties of hyperspectral image is made, and further discusses about it’s specialties. A novel Conditional Random Field (CRF) model is in the field of hyperspectral image classification, and the following efforts are made:(1) Using CRF model to handle the hyperspectral image classification problem. In view of the high dimension hyperspectral image characteristics and nature, employing the corresponding probability model and solving mechanism. This model not only embodies the spectrum correlation of hyperspectral image, it also reflect the space correlation of hyperspectral image, and the implementation simple, efficient.(2) According to the spatial correlation of hyperspectral image, a kind of judge mechanism is put forward, so as to avoid some unnecessary information updates. The CRF model made full use of the spectrum correlation and spatial correlation of hyperspectral image, but in some cases, contextual information is not necessary, this paper introduces a method of minimum distance for limit information renewal. The experimental results show that this method increases the classification accuracy.(3) At present, for hyperspectral image classification and features extraction, supervised methods have relatively good effect. The supervised methods needs extremely rich training sample, which with mark of land-cover category. But many times, it is not easy to get marked category information. We using the conception of convex geometry of hyperspectral image, and extending a kind of sample selection method, to get more available information. The experimental results show that our sample augment CRF is a feasible strategy.