Recursive contextual classification using a spatial stochastic model

摘要

In the past decade, there have been numerous attempts to develop systems for automatic interpretation of digital image data. None of the systems developed for this purpose have made extensive use of context information. Since even manual interpretation of isolated point or area targets is difficult without the use of context, a machine which does not use context has a fundamental limitation.In the course of using contextual information, the first task is to present a stationary stochastic process on a two-dimensional plane. This process is then used as a model, such that correlations between any pair of image cells can be extracted. The model is characterized by a spatial correlation parameter. A flexible coding technique is presented by which the spatial correlation parameter can be estimated. From the coded patterns used for estimating the spatial correlation parameter, a recursive contextual classification procedure is proposed. Some modifications and extensions of the model are specifically developed or substantially refined during this investigation to cover more general situations.Finally, extensive experimental results with remotely sensed multispectral scanner data using the developed model for contextual classification are reported. Both single-stage and recursive contextual classification procedures are tested on real data. The classification results do show the effectiveness and efficiency of the developed contextual classifier.