Quantifying architectural color Quality: A Machine learning integrated framework driven by quantitative color metrics

Abstract

The assessment and prediction of architectural color quality play a pivotal role in urban spatial environments, influencing the aesthetic and psychological dimensions of urban space. Despite the criticality, there exists a lacuna in efficient methodologies capable of quantitatively evaluating architectural color quality. This study endeavors to bridge this gap by introducing a novel framework employing a machine learning approach in conjunction with operationalizable color feature templates for predicting architectural color quality. Four machine learning models - XGBoost, ANN, SVM, and LGBM - are utilized to assess the color quality based on selected color feature indices. Moreover, this research employs SHAP values to elucidate the contribution of various color features towards model prediction. The findings reveal that among the tested models, XGBoost outshines in terms of prediction accuracy. Significant color features including building height, lightness and saturation of primary colors, and red values in both primary and secondary colors were found to exert substantial influence on the model's predictive capacity. This pioneering approach provides a scalable and quantifiable means to evaluate and predict architectural color quality, which has the potential to significantly contribute to urban color planning and evaluation, thereby propelling forward the domain of architectural color research.