Comparison of pixel and object-based image classification based on very high spatial resolution UAV-borne RGB imagery - Baroch case study

Abstract

This case study compares supervised pixel-based and object-based image classification approaches applied to very high spatial resolution RGB imagery acquired by an unmanned aerial vehicle (UAV) for land cover classification. The analysis focuses on the Baroch Nature Reserve in the Pardubice region of the Czech Republic, using RGB images captured by a DJI Mavic 2 DUAL Enterprise UAV at a spatial resolution of 2 cm per pixel. Two classification methods available in ArcGIS Pro, Maximum Likelihood (ML) and Support Vector Machine (SVM), were applied using both pixel-based and object-based classification techniques. Four land cover classes were distinguished: high vegetation, low vegetation, bare soil, and shadows. Classification accuracy was evaluated using 1,000 randomly distributed validation points, with performance quantified by the Kappa coefficient. The results indicate that pixel-based classification achieves higher accuracy than object-based classification, particularly for vegetated areas. These findings suggest that pixel-based approaches are more suitable for high spatial resolution UAV imagery when classifying vegetation-rich environments.