Bridging Long-Term Landsat Time Series and Sentinel-2 Machine Learning to Decipher the Urban–Vacant Land Paradox in Arid Metropolitan Systems

Gulf cities contain extensive undeveloped land inside their urban boundaries, yet its extent is poorly documented. This study examined the urban–vacant land paradox in the Dammam Metropolitan Area, Saudi Arabia, using Google Earth Engine. Annual Landsat composites (2000–2024) were classified with a spectral index rule capturing a built-and-bright-bare (BBB) signal and potential vacant land, and a Random Forest with 14 predictors was applied to 2024 Sentinel-2 imagery. The BBB extent rose from 89,893 ha in 2000 to 94,043 ha in 2015 and fell to 87,875 ha in 2024. The decline was significant but matched the effect of a 0.01 shift in the NDBI threshold and coincided with the withdrawal of Landsat 7, whose composites yield higher NDBI values than those of Landsat 8; it is not interpreted as land-cover change. The negative BBB–vacant correlation (r = −0.864) arose from compositional closure. The 2024 classification mapped 46,766 ha of urban and 42,474 ha of vacant/bare land (overall accuracy 75.05%, κ = 0.620, against WorldCover-derived labels; area-weighted accuracy 72.27%). Accuracy was weakly sensitive to hyperparameters, and disagreement with the reference clustered along class boundaries. Undeveloped land rivals the urban footprint in extent, a finding relevant to infill-oriented planning in arid cities.