Mapping Localized Patterns of Classification Accuracies Through Incorporating Image Segmentation

Land use land cover (LULC) maps are essential for numerous applications, such as urban growth analysis, deforestation, etc. The accuracy of these LULC maps is often assessed using global indicators, and its spatial variations are neglected. To address this issue, this letter proposes to examine local LULC classification accuracy through incorporating a polygon system derived from image segmentation techniques. In particular, LULC classification maps were produced using three widely applied remote sensing classification techniques, maximum likelihood classifier (MLC), artificial neural network (ANN), and random forests (RFs). Then, a polygon system was derived using image segmentation techniques to mitigate intrapolygon variations and enhance interpolygon variations. Finally, a localized LULC classification accuracy map was generated using 2500 randomly selected samples. The derived accuracy maps provide a significant amount of information, with accuracy varying remarkably from polygon to polygon (i.e., from 50% to 100%). Moreover, when the three LULC classification accuracy maps with MLC, ANN, and RF were compared, similar spatial variation patterns have been discerned, indicating the existence of site specific factors that impact classification accuracy. This letter suggests that the developed local LULC classification accuracy maps may serve as a better alternative for numerical accuracy assessment, as well as provide a starting point for further improvements of LULC maps.