FOCUS on Contamination: Hydrology-Informed Noise-Aware Learning for Geospatial PFAS Mapping

Technical Deep Dive: FOCUS on Contamination

Overview

FOCUS is a geospatial deep learning framework for mapping per- and polyfluoroalkyl substances (PFAS) contamination in surface waters. PFAS are persistent environmental contaminants with significant public health impacts, but large-scale monitoring is limited by the high cost and logistical challenges of field sampling. FOCUS integrates sparse PFAS observations with large-scale geospatial data on land cover, hydrology, and industrial activity to generate robust, scalable contamination predictions.

Problem & Context

• PFAS contamination is a major environmental and public health issue, with 97% of Americans exhibiting detectable PFAS levels in their blood
• Measuring PFAS concentrations is expensive, resulting in sparse and unevenly distributed ground-truth data
• This data scarcity leaves substantial uncertainty in identifying contamination hotspots and prioritizing targeted remediation
• Existing modeling approaches, such as hydrological simulations and geostatistical interpolation, have limitations in capturing the complex transport dynamics and uncertainties of PFAS

Methodology

• FOCUS frames PFAS mapping as a geospatial deep learning task, operating directly on multi-channel raster images to preserve spatial dependencies
• The model is trained using a noise-aware loss function that integrates physically-informed pixel correctness priors, such as proximity to industrial dischargers, land cover, and hydrological flow
• These priors help the model learn robust predictions from sparse, noisy labels, while maintaining spatial coherence and scalability

Data & Experimental Setup

• Training and testing data include both fish tissue and surface water PFAS measurements from government agencies and community sampling initiatives
• Raster images are generated around each sample point, integrating land cover, industrial discharger locations, and hydrological flow data
• The model is evaluated on multiple years (2008, 2019, 2022) to assess its ability to fill spatial data gaps

Results

• FOCUS consistently outperforms baseline approaches, including pollutant transport simulations, geostatistical interpolation, and tabular ML models
• It achieves a favorable trade-off between precision and recall, which is critical for robust large-scale environmental screening
• FOCUS is also computationally efficient, reducing feature extraction time from days to hours compared to tabular ML baselines
• Real-world validation on newly collected samples demonstrates the model's ability to generalize beyond the training regions

Interpretation

• FOCUS enables more comprehensive, nationwide perspectives on PFAS contamination risk, complementing existing, state-level monitoring programs
• The model's output can support scientific exploration and hypothesis generation by highlighting regions with unexplained contamination that warrant targeted follow-up investigation
• Integrating domain-specific knowledge into the model's training process helps address the challenges of sparse and noisy environmental data

Limitations & Uncertainties

• FOCUS performance is ultimately bounded by the availability and quality of underlying data
• The current model does not yet attribute predicted risk to specific PFAS compounds and sources, though this is an area of ongoing work
• Predictions should be interpreted as screening-level risk signals to prioritize follow-up sampling, rather than definitive contamination assessments

What Comes Next

• Future work will explore predictive uncertainty quantification to guide targeted sampling and enhance prediction robustness
• Temporal modeling using multi-year data will be investigated to track and forecast evolving PFAS patterns as longitudinal measurements become denser
• The team plans to build on public contamination maps by responsibly deploying the FOCUS framework to support decision-making and environmental justice efforts.