Comparing Kernel Density Estimation and Dbscan For Central Region Detection In Urban Last-Mile Delivery

Segmenting the delivery region is a critical step for simplifying and improving the efficiency of Vehicle Routing Problem (VRP) solutions in last-mile logistics. This paper presents a comparative study of two density-based methods, Kernel Density Estimation (KDE) and DBSCAN clustering, for extracting central regions from large-scale, real-world delivery datasets. Using the Loggi Benchmark for Urban Deliveries (LoggiBUD), which covers diverse Brazilian cities with varying delivery densities, we systematically evaluate the spatial characteristics and robustness of each method. KDE produces smooth and contiguous central regions that capture the overall structure of high-density delivery zones, while DBSCAN isolates compact clusters, demonstrating greater sensitivity to local density variations. Quantitative analysis across multiple regions shows that KDE consistently identifies larger and more realistic central regions, whereas DBSCAN excels at pinpointing ultra-dense delivery hotspots but may fragment the urban center in heterogeneous contexts. Statistical tests confirm significant differences in the area and compactness of regions identified by each method. Our findings provide actionable guidance for researchers and practitioners in urban logistics, highlighting the trade-offs between holistic and localized central region extraction. The results support the use of KDE for comprehensive urban center segmentation and DBSCAN for targeted hotspot detection, with implications for scalable and adaptive VRP planning.