Xinyu Ma

Blue–green infrastructure (BGI) is an important nature-based solution for mitigating urban overheating, and machine learning (ML) techniques offer strong potential in capturing the complex, nonlinear relationships between influencing factors and BGI cooling effects. This study presents a systematic review of 54 journal articles published between 2015 and 2025 that applied ML to assess the cooling performance of BGI. We classified BGI into seven types across three spatial scales, evaluated the metrics and indicators used to characterize the BGI cooling effects, and examined the ML algorithms, model performance, and data sources adopted in reviewed studies. We found that 2D BGI morphology metrics and BGI physical and biological metrics are most frequently used as model inputs, whereas 3D morphology metrics are used less. Surface temperature-related indicators dominate model outputs, while air temperature and human heat stress indicators, which better reflect heat health hazards, remain underutilized. In terms of methodology, ML applications are primarily based on artificial neural networks for temporal regression and tree ensemble algorithms for spatial regression. Building on these findings, we identify key research gaps and advocate that future research should develop comprehensive multi-source and multi-scale BGI databases, improve model transparency and generalizability, and integrate energy-balance information into BGI-ML studies. Finally, we propose a systematic BGI-ML modeling framework to guide future research in this field.

Blue-green infrastructure (BGI) is an important nature-based solution to mitigate the urban heat island effect. Machine learning (ML) techniques can integrate multi-source data to more quickly and accurately assess the cooling effect of BGI and understand the complex nonlinear cooling mechanism. This study systematically reviews 38 journal articles published from 2012 to today that applied ML to evaluate the cooling effect of BGI covering different scales and types. Studies are categorized according to six criteria: Spatiotemporal Scale; ML algorithm; BGI Type; Cooling Effect Metrics; Input Data and Data Source. Artificial Neural Network and Random Forest are the two mostly-used ML algorithms. BGI types include trees, green roof, green wall, water body, open green space and green patches, among which green patches are the most studied while trees and green walls are the least. The background climate, urban morphology and evaluation methods of BGI cooling effect vary significantly among studies and it is thus difficult to compare and synthesize the cooling efficacy performance: Studies that cover multiple climatic zones with long time series meteorological data (e.g., multi-year observation) would support generalizability but these studies are lacking. BGI land surface coverage are the most frequently-used indicators, whereas 3-Dimensional indicators are much less adopted: less than 20% of studies used 3D building indicators and only 10% used 3D vegetation indicators. More than half of the studies’ labeled data source is from remote sensing, while fewer used air temperature and thermal comfort data that could more accurately reflect local and microscale climates. The review underscores the need for more comprehensive BGI studies, covering larger spatiotemporal ranges and more representative datasets to improve the effectiveness and generalizability of BGI-cooling ML models.