Formatted Title
Towards Improved Hydrocarbon Soil Assessment: The Application of Mid-Infrared Spectroscopy and Binary Classification Techniques
Background/Objectives
Environmental contamination challenges, especially those related to the detection, assessment, and remediation of hydrocarbon-contaminated soils, can benefit from more efficient and rapid assessment tools. Traditional laboratory testing methods, while comprehensive, involve extended turn-around times, and can become costly with large numbers of samples. In contrast, in-field extraction kits, which are reliant on solvents and other consumables, demand continual recalibration and can be cumbersome for in-field use.
We are developing a tool which enables on-site, prompt assessment of hydrocarbon contamination in soils, specifically targeting contamination levels above or below the widely recognized regulatory standard of 1000 ppm (0.1% contamination). This tool was evaluated on historical data from multiple test sites, notably in varied environments such as Sumatra, Victoria's coastal wilderness, a French industrial site, and Antarctica. Selected due to their history of soil contamination from projects ranging from small-scale to multi-year large-scale, these sites presented unique challenges, including diverse hydrocarbon contaminants, soil types, and interfering components such as organic matter and carbonates.
Approach/Activities
Our research leverages mid-infrared (mid-IR) spectroscopy using the RemScan, a portable spectrometer. This device utilizes diffuse reflection from soil samples to probe the vibrational modes of hydrocarbons. In essence, the strength of the signatures in the mid-IR region is related to the level of hydrocarbon contamination in the soil.
Our training dataset comprised over 15,000 meticulously prepared calibration standards, spanning a large variety of soils types and spiked with specific diesel concentrations. The industry-adopted threshold of 0.1% contamination served as our benchmark for soil cleanliness. Our methodology involved experimenting with seven different classifiers and various pre-processing and sampling techniques. A significant challenge was differentiating between soil organic carbon (SOC) and total petroleum hydrocarbons (TPH) due to their overlapping signatures, emphasising the importance of our ongoing efforts to refine the differentiation process.
Results/Lessons Learned
The best classifier demonstrated over 90% accuracy in our offline testing on both calibration and validation datasets using real-world data. Notably, superior training performance did not consistently translate to an equivalent validation success. Factors such as the presence of calcium carbonate and high soil organic carbon concentrations proved influential in the classification process.
We are currently working on advancing methods to discern low-level TPH contamination from organic carbon within contaminated samples. Given the challenges with high carbonate soils, there is an ongoing initiative to engineer a specialised classifier to address these samples. It is important to note that the results cited are from offline validations. Our current trajectory involves integrating this software solution into RemScan and performing small-scale, in-field testing with the instrument. These findings will be presented.