Measuring surfaces for Nature-based Solutions in the urban environment: A deep learning approach for Vertical Area Analysis

Abstract

This study aimed to identify and measure potential sites for vertical gardens in urban areas using street view imagery, computer vision techniques, and geospatial analysis. The research addresses the growing need for nature-based solutions in cities to combat the urban heat island phenomenon. A multi-step methodology was developed with data retrieval from Google Street View API, image segmentation using the Grounded Segment Anything Model, and trigonometric calculations for real-world scale conversion. Applied on a study area of 2.22 km² in Zwolle, Netherlands, selected for representative urban characteristics. Results demonstrate variable effectiveness in accurately estimating building dimensions, with best results for structures in close proximity to the viewpoint. Analysis revealed a positive correlation between distance and estimation error, with Mean Absolute Percentage Errors (MAPE) ranging from 7.23% to 14.98% across distances. The approach shows promise in automating the identification of suitable facades for vertical gardens. However, computational constraints and decreasing accuracy at greater distances should be addressed in future research.