Multimodal Data Fusion and Deep Learning for Predicting Air Pollution at Construction Sites

The construction sector is a major contributor to air pollution, emitting large volumes of particulate matter that pose significant health and environmental risks. Construction activities, such as excavation, material transport, and vehicle operation, are key sources of these emissions. While regulatory frameworks establish emission limits, proactive air pollution management at construction sites remains challenging due to the lack of real-time predictive analytics. Existing methods rely on laboratory tests and predefined emission inventories but fail to integrate multimodal data, such as environmental conditions and construction activity information, to improve predictive accuracy. Additionally, they cannot dynamically capture temporal variations in these data. To address these gaps, this paper presents a data-driven system that integrates multimodal data and deep sequential models to forecast PM emissions at construction sites. The proposed system consists of (1) a smart sensor system for multimodal data collection, (2) a fine-tuned you only look once (YOLO) model for detecting construction vehicles and workers, and (3) a deep sequential model that leverages data fusion and long short-term (LSTM) memory networks for air pollution prediction. The system was evaluated through experiments at a selected construction site, including a comprehensive analysis of prediction performance, comparisons with baseline machine learning models, and an ablation study on temporal sequence length and data modality. The proposed system achieved a mean squared error of 0.0034, mean absolute error of 0.0416, and R2 of 0.9590, outperforming baseline models and demonstrating good predictive accuracy.