Comparison of machine learning algorithms for green view index (GVI) prediction using NDVI and urban form metrics

Abstract

Street-level greenery has been widely recognized as an important environmental exposure beneficial for human health. Despite the widespread use of street view images in evaluating street-level greenery, the assessment is constrained by incomplete coverage of street view. This study develops different machine learning (ML) models, including least absolute shrinkage and selection operator (LASSO), support vector machine (SVM), random forest (RF), extreme gradient boosting (XGB), light gradient boosting machine (LGBM) and neural network (NN), to estimate green view index (GVI) along the Swiss Basel-Stadt and Basel-Landschaft cantonal road networks by combining satellite-derived greenery indices and urban form metrics. Models were developed for varying buffer radii (15 m and 25 m) around each street view sampling point. Except for LASSO, the performance of the ML models ranged from moderately strong to strong in explaining GVI prediction (R2 > 0.50) for both random and block cross validation, with the RF model performing best on the test dataset (R2 = 0.75). Multiple feature importance analyses indicate that normalized difference vegetation index (NDVI), leaf area index (LAI), viewshed greenness visibility index (VGVI), average vegetation height (Ave_VH), average building height (Ave_BH), and street aspect ratio (SAR) are the primary factors influencing GVI, and models focusing on rural typology reveal the crucial roles of building coverage rate (BCR), and number of buildings (NoB) in GVI prediction. This study demonstrates that ML is a promising approach for supplementing GVI assessment in the absence of street view images, and provides a comprehensive understanding of how detailed urban form characteristics affect street-level greenery visibility. These findings also offer practical insights for urban greenery management.