Formatted Title:

Seeing the Trees from the Forest: Distinguishing Site Impacts from Urban Background Using Machine Learning

Background/Objectives:

Distinguishing site impacts from background is critical to developing remedial strategies within urban waterways. For many sediment cleanup sites, developing a remedial footprint is accomplished by comparing sediment concentrations to a cleanup value determined from background or reference concentration data. In urban waterways, establishing sediment cleanup values can be complicated by a variety of factors and processes including complex hydrodynamics, historical and current inputs to the system, and resuspension and redistribution of sediment impacts. For sites where inorganics are the primary contaminants, identifying and delineating site impacts are further complicated as metals are ubiquitous in urban environments as well as naturally occurring. As such, development of a multiple lines of evidence framework to identify and quantify site-related inorganic impacts from urban and natural background is a critical component of the remedial strategy. A comprehensive framework approach leveraging multi-variate statistics and machine-learning algorithms was developed to assess sediment background and site-related impacts in an urban waterway in the eastern US and guide the site management strategy.

Approach/Activities:

Application of machine-learning algorithms capable of developing chemical fingerprints for various sources enables new approaches to distinguish site-related constituents from urban background. In this case study, advanced, multi-variate statistical analyses including gaussian mixture modeling (GMM) and linear discriminant analysis (LDA) were applied to a large public dataset to identify and distinguish sample locations characteristic of anthropogenic and natural background from samples impacted by potential contamination sources within the waterway. The identified background samples provided an estimate of background concentrations of key site constituents. Additionally, self-training machine-learning fingerprinting algorithms (alternating least squares [ALS] and non-negative matrix factorization [NMF]) were applied to chemical samples from site sediment cores to distinguish site-related constituents from urban background and develop a fingerprint of these impacts.

Results/Lessons Learned:

Multi-variate statistical analyses of the publicly available dataset identified that regional sediment samples collected within the ~10 mile waterway aligned with four categories: two groupings representative of differing levels of contamination (moderate-high metals concentrations and low-moderate metals concentration), one grouping characteristic of diffuse, anthropogenic impact, and one grouping representative of natural background conditions. This unsupervised and unbiased approach enabled efficient assessment to remove samples associated with contamination and estimate appropriate background concentrations based on the anthropogenic and natural background data clusters. Application of machine-learning algorithms to the site data identified fingerprints for four end members including two signatures for site sources, a natural background signature, and an anthropogenic background signature. The contribution of each source to each sample identified locations and depths dominated by the site source signatures from which delineation and the potential remedial footprint were derived. Machine-learning applications to assess sediment data, including at sites with complex source and background mixtures, have the potential to shift the historical site assessment and management paradigms by enabling efficient characterization of background, site contributions, and delineation of site impacts to focus and guide the site management strategy.

Steve Finn

Affiliation

Track

Refine Results

Seeing the Trees from the Forest: Distinguishing Site Impacts from Urban Background Using Machine Learning