Work

Assessing airflow perturbations in naturally ventilated barns: A comparative study of image similarity metrics using machine learning

Year: 2025

Type: article

Abstract: While natural ventilation is crucial for maintaining an ideally optimized air exchange rate (AER) and animal wellbeing, measuring the accurate airflow patterns is indeed challenging. This study investigates the behavior of a wide range of image similarity indices to detect differences between computational fluid dynamics (CFD) images of barn ventilation. To this end, a comprehensive dataset of barn airflow patterns was generated via CFD simulations under diverse conditions, then systematically distorted using swirl transformations to simulate controlled perturbations in airflow images with known ground truth. A broad set of image similarity indices was included to evaluate their sensitivity in the context of CFD images. We further fitted 14 state-of-the-art machine learning regressor models to predict the ground truth perturbation from the similarity indices and investigated which index has the highest feature importance. The strongest predictive performance (R 2 = 0.89) is achieved using a stacked model of top 6 regressors, namely LightGBM, XGBoost, CatBoost, Gradient Boosting, Extra Trees and Random Forest, achieving R 2 of 0.879, 0.872, 0.870, 0.868, 0.860 and 0.840, respectively. We also show that DCT subband similarity (DSS), Mean Deviation Similarity Index (MDSI) and Deep Image Structure and Texture Similarity (DISTS) have the highest feature importance in predicting the swirl extent. Our findings further reveal that top 7 image similarity metrics can reasonably predict the perturbations in CFD simulation images. Lastly, we validated the trained models on a real-world low resolution dataset from a dairy barn, showing a high empirical power of 0.96 for our stacked model.

Source: Computers and Electronics in Agriculture

DOI: 10.1016/j.compag.2025.111186