Abstract. This study investigates the behaviour of segmentation accuracy metrics in object-based image analysis (OBIA) using synthetic urban objects. Monitoring land cover through high-resolution imagery relies heavily on accurate segmentation, which directly influences classification performance. A synthetic dataset was created with a fixed-size reference square and varying-sized square segments positioned systematically to analyse spatial and geometric relationships. Six widely used accuracy metrics were evaluated: Area Fit Index (AFI), Match (M), Quality Rate (QR), Over-Segmentation (OS), Under-Segmentation (US), and Quality of Object Location (qLoc), representing both area-based and location-based criteria. Results reveal that area-based metrics generally show consistent trends and similar sensitivity to changes in segment size and geometry, while location-based metrics exhibit independent patterns emphasizing spatial positioning and locational accuracy. This divergence highlights the limitations of relying solely on either metric type, advocating for an integrated evaluation framework combining both area and location criteria to achieve a more comprehensive assessment of segmentation quality. The study suggests that future research should incorporate more complex and irregular urban object shapes and explore additional metrics, such as boundary-based or context-aware measures. Furthermore, the identification of optimal segmentation configurations guided by these metrics could enhance training data quality for deep learning applications in urban object classification.
Avsar et al. (Tue,) studied this question.