Classification of dune vegetation from remotely sensed hyperspectral images

Vegetation along coastlines is important to survey because of its biological value with respect to the conservation of nature, but also for security reasons since it forms a natural seawall. This paper studies the potential of airborne hyperspectral images to serve both objectives, applied to the Belgian coastline. Here, the aim is to build vegetation maps using automatic classification. A linear multiclass classifier is applied using the reflectance spectral bands as features. This classifier generates posterior class probabilities. Generally, in classification the class with maximum posterior value would be assigned to the pixel. In this paper, a new procedure is proposed for spatial classification smoothing. This procedure takes into account spatial information by letting the decision depend on the posterior probabilities of the neighboring pixels. This is shown to render smoother classification images and to decrease the classification error.