<![CDATA[GeoAI technology, which integrates artificial intelligence with spatial analysis, offers a novel approach to extracting urban object information from high-resolution imagery. This study applies Mask R-CNN with a ResNet-50 backbone architecture to detect vehicle objects in orthophoto imagery derived from the processing of 100 UAV photographs over an urban area in Switzerland. A total of 80 vehicle objects were annotated and partitioned into training (70%), validation (15%), and testing (15%) datasets. Model evaluation was conducted using a multi-threshold Intersection over Union (IoU) approach at values of ≥0.5, ≥0.75, and ≥0.95, and analyzed through a confusion matrix alongside Precision, Recall, F1-score, and Mean Average Precision (mAP) metrics. The results demonstrate that the model achieved Precision and Recall scores of 1.00 at IoU ≥0.5; however, performance declined at stricter thresholds, with an aggregate mAP of 0.56, indicating moderate overall performance. These findings suggest that the model is effective for macro-spatial analytical needs such as vehicle count estimation and distribution mapping, yet remains insufficiently stable for applications requiring high geometric precision. Conceptually, this study underscores the importance of multi-threshold evaluation in the application of deep learning for urban spatial analysis, while demonstrating the potential of GeoAI integration in data-driven urban planning.]]>
Copyrights © 2026
