AI Agentic Workflow for Sinkhole Risk Prediction Near Underground Assets

This workflow automates the detection of millimeter-scale ground movement using InSAR satellite data before a sinkhole forms, shifting from a model of emergency response to a scheduled, high-priority inspection regime. By predicting failure points along pipelines or near foundations, it prevents ruptures, service outages, and environmental incidents that typically cost millions in remediation, fines, and lost revenue.

Manual correlation of subsidence data with geological maps and asset registers is slow and error-prone. This workflow uses orchestrated agents to fuse InSAR displacement time series, karst geology layers, and pipeline GIS data, automatically scoring risk and generating a prioritized dig-site list. It eliminates low-value inspection targets, ensuring geotechnical crews and expensive direct assessment tools (e.g., ground-penetrating radar) are deployed only where the risk is highest.

Sinkhole risk is translated from a geospatial model into a financial and operational plan. The workflow automatically generates capital project requests for ground stabilization, pipeline rerouting, or proactive reinforcement, complete with cost estimates and recommended timelines based on risk velocity. This creates a defensible, data-driven pipeline for capital allocation, moving budget from reactive emergency funds to planned, lower-cost preventive investments.

For pipelines in sensitive areas, demonstrating proactive geohazard management is a regulatory imperative. The workflow automates the assembly of evidence packages—including displacement charts, risk scores, and inspection schedules—for submissions to agencies like PHMSA. It also triggers automated landowner notifications for planned investigations, reducing community relations friction and creating a full audit trail for compliance.

The system is designed for continuous learning. As new InSAR data arrives and ground-truth results from dig sites are logged, agents automatically retrain and calibrate the risk prediction models. This creates a self-improving feedback loop where the accuracy of sinkhole forecasts increases over time, directly lowering the cost of false positives and increasing confidence in the operational schedule.

Implementation is built on a LangGraph-orchestrated agent system that ingests satellite feeds, queries geological databases (e.g., USGS), and integrates with pipeline integrity management systems (PIMS) like GE Smallworld or ESRI ArcGIS. The workflow includes mandatory human-in-the-loop approval gates for high-cost mitigation actions and a centralized observability dashboard for tracking risk scores, work order status, and model performance across the asset network.

This core agent ingests and processes Interferometric Synthetic Aperture Radar (InSAR) satellite data to detect millimeter-scale ground subsidence over time. It correlates displacement vectors with geological maps of karst terrain and historical mine workings to calculate initial sinkhole formation probability. The agent automates what was a manual, expert-driven analysis of time-series geospatial data, turning raw satellite feeds into actionable risk signals.

This connector agent queries the enterprise Geographic Information System (GIS) and Pipeline Integrity Management System (PIMS) to map subsidence risk zones against the precise location, depth, and criticality of buried pipelines, fiber optic cables, or foundation pilings. It enriches the geohazard model with asset-specific consequence scores, automating the manual cross-referencing that previously delayed risk assessment.

The central orchestrator (built with frameworks like LangGraph) combines the probability output from the modeling agent with the consequence data from the GIS integrator. It applies business rules to generate a unified, ranked risk score for each asset segment. This component automates the prioritization logic, ensuring field resources are dispatched to the highest-probability, highest-consequence locations first, directly improving capital allocation.

This agent translates high-risk predictions into operational action. It automatically creates fully populated inspection or mitigation work orders in systems like SAP, IBM Maximo, or Oracle with recommended actions, required equipment, and linked geospatial data. It eliminates the administrative lag and data re-entry between risk identification and field scheduling, closing the automation loop from detection to dispatch.

A critical control layer for high-cost mitigation actions (e.g., ground-penetrating radar surveys, grouting programs). This gateway routes work orders exceeding a cost or risk threshold to a geotechnical engineer for review and approval via a structured dashboard. It ensures expert oversight is preserved for major decisions while letting low-risk validations proceed automatically, balancing speed with governance.

This component tracks the performance of the entire workflow. It logs prediction confidence scores, correlates them with subsequent field inspection findings, and feeds discrepancies back to retrain the geohazard model. This creates a closed-loop, self-improving system that increases prediction accuracy over time, turning field data into a continuous feedback signal for the AI agents.