Typology-Based Verification of Green-Roof Potential Sites in Lisbon through WorldView-2, Sentinel-2, and Building-Objects Validation

Abstract

Green-roof inventories for municipalities should be able to discriminate between confirmed roofs greening and vegetation intersecting building footprints. It is true especially for Lisbon with historical red clay roofs, flat individual roofs, trees on the borders of streets, and communal podium-type green spaces in a dense urban fabric. This paper investigates whether Lisbon’s candidate inventory reflects the condition which should be prioritized for field validation when the accuracy of image-class classification is high but the object assignment differs by morphology. The dataset consists of an eight-band WorldView-2 image obtained on 18 February 2022, twelve observations made monthly with Sentinel-2 in 2023, 54,800 municipality building footprints, 4,015 grouped-building detections, 371 final candidates, and 196 verified green roofs. The investigation follows five-stage procedure that is built into the Lisbon record: class-specific reliability of the image classification, successive elimination, beta-binomial posterior credibility, seasonal NDVI margins, and inspection burden distribution. The classification audit of the WorldView-2 image is quite successful as evidenced by the overall accuracy value of 0.950, Cohen’s κ of 0.914, vegetation F1 of 0.960, and unsuitable-roof F1 of 0.947. These numbers confirm the greenness and materials-based screening but not the roof attribution. Validation of the objects is very typologically specific: 50 out of 176 individual-roof candidates pass the test, versus 146 out of 195 shared space candidates. The verification odds for the shared space path are 7.51 higher than for the individual-building path, and the reliability-adjusted posterior intervals stay well-separated. The seasonal NDVI discriminates the two types too, since top-level roofs span 0.25-0.31 margins around the 0.30 gate, while shared space roofs are located above the gate at 0.32-0.49. In summary, Lisbon’s candidate map includes two separate streams of verification: individual roofs need specific error detection while common space roofs present the high yielding approach to completion of the city’s inventory.