AI & Emerging Technologies

Large infrastructure projects generate thousands of environmental commitments across permitting, environmental review, and construction phases. Translating those commitments from dense regulatory documents into a trackable, auditable system of record required intensive manual effort — copying text from PDFs, manually categorizing obligations, and managing revisions over multi-year project timelines.

We embedded an AI-assisted commitment parsing capability into Beacon, ESA’s compliance platform. The system reads environmental approval documents and uses LLM-based extraction to pull commitment titles, identifiers, and timing requirements, structuring them into searchable, trackable records. Users review and confirm AI-generated suggestions, which retain links back to source documents. The result is a transparent, auditable workflow that keeps experienced professionals in control.

Decades of critical groundwater data were locked in thousands of scanned Well Completion Reports — semi-structured documents spanning 80 years in varied formats, including handwritten records. The data was essential for basin-scale groundwater modeling and California Sustainable Groundwater Management Act (SGMA) compliance, but the cost and time of manual extraction was not feasible.

For the Merced Irrigation-Urban Groundwater Sustainability Agency, we built a cloud-based machine-learning pipeline on Microsoft Azure Document Intelligence Studio to extract and geolocate well addresses, parcel numbers, and lithology descriptions from over 10,000 Well Completion Reports (WCR). Building on those results, we expanded the approach for the Yolo Subbasin Groundwater Agency — refining classification and model training to extract lithologic and well construction data from approximately 13,000 additional records. The outcome is a reusable, documented Python pipeline applicable across California basins with legacy WCR archives.

Water Quality Management Plans (WQMPs) are the backbone of stormwater compliance — but each one can run hundreds of pages of dense technical language. Agencies managing large WQMP portfolios had no efficient way to extract comparable structured data across documents. Manual entry was slow, inconsistent, and difficult to audit.

Using the OpenAI API, we developed an extraction workflow combining prompt engineering, JSON schema mapping, and validation routines to automatically pull key WQMP attributes — site characteristics, treatment types, Best Management Practices (BMP) selections — directly from PDF files. The system assigns confidence scores and cites the source page and paragraph for each extracted data point. Client reviewers verify results within the platform interface, maintaining accuracy and verifiability.

Large-scale environmental monitoring programs including habitat assessments, land cover surveys, and vegetation mapping have historically required extensive manual interpretation of aerial and satellite imagery. At watershed or corridor scale, manual methods are too slow and inconsistent to support near-real-time environmental management decisions.

ESA’s geospatial team applies machine learning and deep learning models including convolutional neural networks trained on multi-spectral satellite imagery, to automate habitat classification, land cover change detection, and vegetation condition analysis. Output is delivered through interactive web applications and GIS environments, enabling managers to track conditions over time, flag anomalies, and prioritize field verification. Our pipeline supports Planet, Sentinel, and Maxar sensor platforms.

ESA’s Groundwater Accounting Platform (GAP) is a powerful tool for California Sustainable Groundwater Management Act (SGMA) compliance data management, but accessing its analytical depth required programming skills that many groundwater scientists and agency staff didn’t have. We wanted to demonstrate that AI could democratize access to complex environmental data tools.

We hosted ESA’s first AI-led coding hackathon — bringing together participants with little or no coding background to interact with the GAP API using Google Colab and Gemini AI as a virtual coding teammate. In two hours, five project teams developed working data applications: correlating evapotranspiration and land subsidence, screening fields for fallowing potential, benchmarking per-parcel water use, and flagging potentially unreported wells.