Research map / Detection and segmentation / Open-set and few-shot detection
Alberta Wells Dataset: Pinpointing Oil and Gas Wells from Satellite Imagery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the Alberta Wells Dataset of ~188k Planet Labs 3m satellite patches with ~213k labeled oil and gas wells (abandoned, suspended, active) and benchmarks segmentation and object detection baselines.
Problem
Many abandoned oil and gas wells are unrecorded and leak methane and toxins; existing remote sensing well datasets are small, region-limited, and mostly contain only active wells, so abandoned/suspended wells are hard to locate at scale.
Approach
A large, expert-verified benchmark of multispectral (RGB+NIR) Planet Labs patches labeled with AER well locations across all well states, with a cluster-based geographic train/val/test split, framed as both binary segmentation and object detection.
Key results
- Best binary segmentation: U-Net with EfficientNetB6 reaches IoU 60.4±0.3 and F1 64.8±0.4; UperNet Swin-small gives the highest recall (73.1±0.1).
- Object detection: DETR achieves the highest mAP50 (15.22±0.28) and mAP50:95 (38.45±0.32), while SSD Lite has the best IoU0.5 (65.07±0.03).
- Adding NIR band improves FCOS mAP50:95 from 20.00 to 30.46 and U-Net IoU from 56.6 to 58.0; training on all well types beats active-only training on IoU (0.576 vs 0.502) on the test set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the NIR band to RGB input (ablation) | U-Net ResNet50 · Alberta Wells test set, binary segmentation | IoU: 56.60$\pm$0.44 -> 58.00$\pm$0.50 | Helped |
| Add the NIR band to RGB input (ablation) | FCOS ResNet50 · Alberta Wells test set, object detection | mAP@50:95: 20.00$\pm$10.40 -> 30.46$\pm$3.11 | Helped |
| Train on all well types instead of active wells only (ablation) | U-Net ResNet50 · Alberta Wells test set, binary segmentation | IoU: 0.502 -> 0.576 | Helped |
| Train on all well types instead of active wells only (ablation) | U-Net ResNet50 · Alberta Wells test set, binary segmentation | Precision: 0.998 -> 0.913 | Hurt |
Surprising findings
- Training only on active wells yields near-perfect precision (0.998) but poor recall (0.502), failing to detect abandoned and suspended wells.
- Object detection mAP50 is very low for all models (at most 15.22), despite reasonable IoU scores, suggesting confidence/region proposal issues with dense wells.
Limitations
- Ground-truth well locations come from AER records, which may miss true wells, adding label noise and possible false negatives.
- Performance degrades in high-density well regions and for abandoned/suspended wells with subtle visual signatures.
- Dataset is restricted to Alberta; transfer to other regions is untested.
Experimental setup
Datasets: Alberta Wells Dataset AER ST37 well registry PlanetScope 4-band imagery
Benchmarks: Alberta Wells Dataset
Metrics: IoU F1 score precision recall IoU at 0.1/0.3/0.5 mAP50 mAP50:95
Baselines: U-Net PAN DeepLabV3+ Segformer UperNet RetinaNet Faster R-CNN FCOS SSD Lite DETR
Builds on: U-Net DETR FCOS Faster R-CNN RetinaNet SegFormer UperNet
Closest papers by meaning
- GLH-Water: A Large-Scale Dataset for Global Surface Water Detection in Large-Size Very-High-Resolution Satellite Imagery (AAAI 2024 · similarity 0.83)
- AllClear: A Comprehensive Dataset and Benchmark for Cloud Removal in Satellite Imagery (NeurIPS 2024 · similarity 0.81)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.80)
- The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf (CVPR 2025 · similarity 0.80)
- Kuro Siwo: 33 billion $m^2$ under the water. A global multi-temporal satellite dataset for rapid flood mapping (NeurIPS 2024 · similarity 0.80)
- Generalizable Disaster Damage Assessment via Change Detection with Vision Foundation Model (AAAI 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.