Abstract
Millions of abandoned oil and gas wells are scattered across the world,leaching methane into the atmosphere and toxic compounds into the groundwater.Many of these locations are unknown, preventing the wells from being pluggedand their polluting effects averted. Remote sensing is a relatively unexploredtool for pinpointing abandoned wells at scale. We introduce the firstlarge-scale benchmark dataset for this problem, leveraging medium-resolutionmulti-spectral satellite imagery from Planet Labs. Our curated datasetcomprises over 213,000 wells (abandoned, suspended, and active) from Alberta, aregion with especially high well density, sourced from the Alberta EnergyRegulator and verified by domain experts. We evaluate baseline algorithms forwell detection and segmentation, showing the promise of computer visionapproaches but also significant room for improvement.