Researchers from the University of Massachusetts presented a machine-learning model designed to identify unsewered areas and approximate whether a parcel is served by sewer or an on-site system. UMass described the model as a statewide gap-filling tool to support the wastewater needs assessment where complete ground-truth data are not available.

UMass explained the model uses parcel-level features (building density, road network, census variables, parcel value, distance to treatment facilities and other spatial inputs) to make predictions in two stages: first, whether a parcel has a need for sanitation infrastructure; second, whether that parcel is likely sewer-served or on-site-served. The model outputs both a classification and a confidence score for each parcel.

Project staff emphasized key caveats: model predictions are “good approximations to the best of our knowledge” and should not be used as regulatory determinations or to publish parcel-level labels. "The predictions that are made by models, they're good approximations to the best of our knowledge," a presenter said. The team also said the model will not publicly display parcel-level data to protect privacy.

UMass said county-level label coverage varies and showed county-by-county prediction confidence; accuracy on held-out test labels ranged across counties but was acceptable in many places (mid-80s to mid-90s in counties with labeled data), allowing staff to infer broader coverage where labels are missing. Advisory-group members urged careful vetting, multiple validation steps with regional boards and counties, and layered public release protocols to avoid prematurely labeling properties.

Next steps: UMass will continue to refine the model and work with regional boards on validation. The advisory group and State Water Board staff said the model will be updated as regional ground-truth data become available and that the model will inform — but not replace — on-the-ground verification and regulatory decisions.