A Comparative Analysis of Machine Learning and Deep Learning for Rooftop Vegetation Identification: Supporting Evidence-Based Urban Governance in Dhaka

Source abstract

Dhaka, one of the world’s most densely populated megacities, has faced a severe ecological decline, with green cover plummeting from 44.80% in 1975 to approximately 24.50% by 2005. In response, urban rooftop farming has emerged as a vital adaptation strategy to mitigate the urban heat island effect and air pollution. Objective: This study evaluates the transition from “pixels to policy” by testing automated identification methods for URF to support evidence-based urban governance, specifically the 10.00% holding tax rebate offered by the Dhaka North City Corporation. Utilizing high-resolution (3 cm) drone imagery across three diverse areas of interest—representing planned, organic, and mixed-use urban fabrics, the research compares the performance of Support Vector Machines, U-Net, and Text-Segment Anything Model. Accuracy was validated using a confusion matrix based on 1000 randomly stratified sample points. The SVM model emerged as the most reliable, achieving a Kappa index of 0.74 and 100.00% user accuracy for identifying rooftop vegetation, significantly outperforming the U-Net model (Kappa 0.14). Spatial analysis quantified a distinct “green divide,” revealing that while planned residential zones achieved over 7.50% rooftop greening coverage, dense organic settlements were limited to 6.00%. The study concludes that high-accuracy SVM-based identification provides a scalable foundation for automating fiscal incentives. To bridge the socio-spatial green divide, policy interventions must shift toward inclusive greening strategies, such as vertical farming, and formal integration of URF into Dhaka’s blue-green infrastructure networks.