Abstract

This abstract describes a novel approach to quantifying dust emissions in mining activities using machine learning models to analyze images captured by direct single-lens reflex (DSLR) cameras such as smartphone and surveillance cameras. The study's primary focus is mining operations, such as hauling, crushing, blasting, and loading, where dust generation is a significant environmental concern.

The core of the research involves developing a machine-learning model capable of identifying and quantifying vehicle-induced dust clouds from images. The model's development was grounded in extensive field experiments on ten unsealed road segments conducted by De Silva, 2023. These segments were selected for their variety in road surface materials and were tested under various climatic conditions to accurately capture the dust clouds generated by vehicles. A crucial part of these experiments was using a direct single-lens reflex (DSLR) camera, which was employed to capture the dust clouds produced by a utility vehicle traveling at different speeds.To complement the visual data, a research-grade dust monitor was used to measure the dust emissions due to traffic. This approach allowed for a robust dataset consisting of approximately 210,000 images. These images were meticulously processed and refined to yield around 7,000 high-quality images. These selected images were then manually annotated to create masks for dust segmentation, a critical step in training the machine-learning model. For the initial evaluation of the model, a truncated sample of approximately 900 images from this dataset was used. The model was based on the U-Net architecture, a well-known framework in image segmentation. The baseline performance of the model on this sample set provided initial insights into its efficacy and potential use for mining operations activities. 

Further model development and testing were conducted using a dataset comprising several hours of surveillance cameras, drone cameras, and smartphone camera images from open pit mines, explicitly focusing on operations like haul roads, blasting sites, loading areas, and crushing facilities. This dataset was contextualized with additional data, such as weather conditions and sunlight, which are known to impact dust visibility. The goal was to create a visible dust monitoring solution adaptable to regular surveillance video cameras, drones, and smartphone images and videos.

The proposed solution offers a 10-scale range of insights into pollution levels, ranging from scenarios with no visible pollution to conditions of extreme air pollution. This spectrum of detection capability makes the solution highly versatile and applicable in various environmental monitoring scenarios.

The solutions enhance the use of surveillance cameras and smartphone images to monitor air quality quantitatively and objectively, complementing and enriching standard EPA dust monitoring devices. The envisioned application of this solution is multifaceted. It includes air quality monitoring, dust control process automation with control bonds, forecasting air quality, and comparing and improving dust control measures over time. By leveraging the capabilities of machine learning and image analysis, this approach offers a significant advancement in proactive and real-time environmental monitoring, particularly in operations where dust emissions are a significant concern.