Optimization of urban land-cover classification workflow based on geographic-object analysis using very-high-resolution imagery

Abstract

A recurring problem in Geographic-Object Based Image Analysis (GEOBIA) is the need to tune each one of the three phases involved in the process, segmentation, feature selection, and classification. This paper presents the optimization of a GEOBIA-based urban land-cover classification workflow using very-high-resolution imagery. Two classification workflows are evaluated: (i) a non-optimized workflow; and (ii) an optimized workflow. In the segmentation and classification phases, both workflows used the multi-resolution segmentation algorithm and the random forest classification algorithm. In addition, important spectral, geomorphometric, and textural features are identified as significant predictor variables for the final classification. It is shown that the classification accuracy of every land-cover category increases with optimization, resulting in an overall accuracy increase of 9.34% compared with no optimization. The results show the substantial impact that optimization has on final classification output and suggest the importance of its adoption as a good practice in GEOBIA-based land-cover classification.