Classification of 10 m-resolution SPOT data using a combined Bayesian Network Classifier-shape adaptive neighborhood method

A hybrid inversion method that combines a Bayesian Network Classifier (BNC) with shape adaptive neighborhoods (SANs) is proposed for the classification of 10-m resolution remote sensing images. BNC uses a directed acyclic graph (DAG) to describe variable relationships. We define feature bands and land use/cover types as the feature and class variables, respectively, and describe them as nodes in the DAG. BNC only uses the posterior probability of the class node, and predicts the class with the highest posterior probability. A SAN, containing spectral, textural, and shape features, is used to study the Bayesian network structure, in contrast to methods which only use spectral features. The classification results of the proposed SAN-BNC are compared with those of the spectral-based Maximum Likelihood Classifier (MLC), the SAN-based MLC, and the SAN-based Support Vector Machine (SVM) using Guangzhou as a case study. Our results show that BNC and SVM have superior inference abilities relative to MLC. Ten meter resolution images will furnish better classification results using the proposed SAN-BNC procedure.