An Improved JSEG-Based Algorithm for Segmentation of Categorical and Remote Sensing Classification Maps

Categorical raster maps derived from remote sensing classifications are widely used to describe land cover, landforms, and other environmental characteristics. However, these products are usually analysed at the pixel level, which limits the identification of larger spatial structures and coherent landscape units. The J-image Segmentation (JSEG) algorithm provides a promising framework for region delineation, but it was originally developed for natural color imagery and relies on fixed scale assumptions that are poorly suited to thematic geospatial data. To address these limitations, we developed GeoJSEG, a modified version of JSEG designed for remote sensing categorized spatial datasets. The method operates directly on classified raster layers, introduces user-defined scale parameters, and employs Jensen–Shannon Divergence during region merging, enabling segmentation that better reflects the spatial organization of geographic phenomena. GeoJSEG was evaluated using synthetic categorical maps, natural RGB images, orthophoto-derived data, land-cover maps, and geomorphon representations of terrain forms. For the synthetic class map, the segmentation quality measure J¯ decreased from 0.088 to 0.012 (better quality), while for orthophoto data, it decreased from 0.059 to 0.025 (better quality). Improvements were also observed for land-cover data and most natural-image datasets. The results demonstrate that GeoJSEG extends JSEG toward scale-aware regionalization of thematic raster data and provides a practical tool for post-classification analysis of remote sensing products.