Hierarchical Classification and Knowledge-Based Decision Tree for Land Use/Land Cover Classification with Very High Resolution Imagery

Abstract

Very high resolution aerial imagery, while providing a very high level of detail of the land surface, introduced new challenges for land use / land cover (LULC) classification.  Finer spatial resolution increases the spectral heterogeneity of the land cover features being classified, and markedly increases the number of shadow pixels.  The result is a very large proportion of confused pixels sprinkled throughout the image which adversely affects the accuracy of LULC classification. This paper describes a hierarchical classification approach using knowledge-based decision tree for LULC classification (coniferous, deciduous, bare ground, water, and roads) with a very-high resolution (30cm) multispectral imagery.  Shadowed or confused pixels were identified and underwent separate processes to effectively reassign those pixels to the base LULC classes, which was determined by the knowledge-based decision tree in this order: roads, water, coniferous, bare ground, and deciduous based on uniqueness of their spectral reflectance and whether they can be clearly separated from other classes.  The overall classification accuracy is 89.4% with a Kappa of 0.85.