An ontology-based multi-class terrain surface classification system for aerial imagery

Automatic classification of terrain surfaces from aerial imagery is essential for military planning, unmanned ground vehicle navigation, environmental monitoring, earth resource management, etc. In this paper we present a terrain surface classification system based on classification ontology to deal with complex multi-class terrain surface identification. The system leverages both humans' a priori knowledge about the characteristics and ambiguity of different terrain classes and a powerful fuzzy decision forest technique to construct an effective, robust, and easily extensible terrain surface classification system. We have tested the developed system on a set of challenging real aerial imagery covering 4000×4000 square meters geospatial areas in California state and achieved 85.5% classification accuracies over eight major terrain classes.