Reliability-Aware Urban Green-Space Detectability from Sentinel-2 and OpenStreetMap Evidence in the Guangdong–Hong Kong–Macao Greater Bay Area

Abstract

The problem of urban green-space mapping in a highly concentrated metropolitan region involves consideration of the reliability of the decision itself rather than just the separation of vegetation from non-vegetation evidence. The Greater Bay Area of Guangdong-Hong Kong-Macao features high-rise neighborhoods, complex impervious surfaces, coastal towns, peri-urban vegetation, and an imbalanced amount of VGI data, indicating that the mere aggregation of accuracy measures will be inadequate for urban planning purposes. In this paper, it was examined whether reliability-weighted conformal graph calibration applied to the Sentinel-2 urban green-space evidence in combination with OpenStreetMap would be capable of discriminating among reliable mapped vegetation, city-level omission, and limitations imposed by structural connectivity. The reliability-weighted conformal graph calibration analysis used such inputs as the OpenStreetMap semantics for vegetation evidence, weight-sensitivity reliability scores, classifier confusion matrix numbers, city-level precision and recall scores, UGS area estimates, landscape metrics, and multi-scale partition analysis. The weighted reliability resulted in 0.8372 of mean accuracy and 0.8267 of mean F1 score, whereas unweighted setting decreased accuracy to 0.818 and F1 to 0.812. The best planning-oriented results of errors were observed for W-SVM with 355 false positives, 0.104 false positive rate, 0.896 specificity, and 0.785 utility. The detectability varied from 97.60 percent in Shenzhen to 67.42 percent in Kaiping, thus clearly showing a distinction between high-confidence cities and fragmentation-sensitive cities with considerable pressure on omission. Finally, the 10 m UGS estimation of 139,427.06 ha appeared to lie between ALCC, ESA WorldCover, and CLCD estimations, indicating substantial disagreements among urban green-space products. The MSPA analysis demonstrated that core and edge classes prevail in the foreground of vegetation evidence, whereas bridge, branch, and loop connectors combined account only for 3.11%.