Alberta Wells Dataset: Pinpointing Oil and Gas Wells from Satellite Imagery

Thank you very much for the thoughtful and constructive review. We appreciate your recognition of the dataset’s scale, quality, and importance for climate-relevant applications, as well as your positive assessment of our methodological rigor and experimental design.

Below, we address your questions in further detail. Please let us know if there are other points you would like us to clarify.

Generalization beyond Alberta

Thank you for raising this point. While we do not claim that models trained on Alberta data will generalize zero-shot to other geographies, we believe Alberta provides a valuable and representative testbed for developing and benchmarking detection models. It is also an especially impactful region in its own right, given its position as one of the largest oil-producing regions globally. That said, we fully agree on the importance of broader applicability, and future iterations of the benchmark will include other regions to test transfer learning, as you suggest.

Impact of AER Record Incompleteness / False Negatives

Thank you for raising this important point. Like many regulatory sources, the Alberta Energy Regulator (AER) dataset is known to have occasional omissions—particularly for older, undocumented, or improperly decommissioned wells. While our dataset reflects the most complete and authoritative ground truth publicly available (from the AER ST37 database), we acknowledge the presence of potential false negatives stemming from unrecorded sites.

That said, one of our core motivations is to support exactly this gap: enabling ML models trained on known wells to help identify undocumented or potentially missing ones. In that sense, false negatives in the label set may represent real-world deployment opportunities for models rather than just noise in the training data. We are currently exploring whether it is possible to examine by hand a subset of “false” well detections from our models to see whether these in fact represent omissions in the dataset. We can endeavor to have this ready for a camera-ready version.

Precision Drop in U-Net (Table 7)

Thank you for noticing this. The drop in precision from 0.998 to 0.913 when training on all well types (compared to only active wells) is due to a tradeoff between generality and specificity. Including all three types (active, suspended, abandoned) broadens the model’s learning target, introducing more subtle and ambiguous visual patterns—especially for older, overgrown, or decommissioned sites. This leads to slightly more false positives in some regions, but substantially improves recall, F1-score, and class-wise generalization, as demonstrated in our class-specific breakdown .

Thank you for your extremely thorough and thoughtful review. We very much appreciate the helpful and constructive feedback. We respond to specific comments and questions below:

Architecture comparisons and hyperparameter tuning:

Thank you for raising this point. We agree that fully tuning hyperparameters in all experiments could be valuable, but believe our use of standard performant settings during training nonetheless provides meaningful benchmarks, especially given that we did not observe significant sensitivity in our models.

We have obtained additional experimental results on the object detection task for the transformer-based model DETR, using a ResNet50 backbone. As shown below, DETR achieves strong localization performance, particularly at higher IoU thresholds and mAP.

Despite these performance gains, our more lightweight models (e.g. SSD Lite and FCOS) may be more usable in practice, especially for remote sensing practitioners outside of ML.

Table R1 : Object detection results on the test set. We report Intersection over Union (IoU) at thresholds 0.1, 0.3, and 0.5, as well as mean Average Precision (mAP) at IoU = 0.5 and IoU ∈ [0.5, 0.95].

Multiclass vs binary problem

Thank you for bringing up this point. We agree that there are multiple different use cases (and relevant stakeholders) associated with detection of abandoned vs active wells. However, binary detection is nonetheless very useful, as (i) determination of well types can often be done by hand, as detection is the more time-intensive step, and (ii) where centralized records for wells exist, active wells are more likely to be documented than abandoned wells, meaning that in many cases newly discovered wells are likely to be abandoned. There is also considerable noise in well status labels, and the visual distinctions between the different classes are blurry (e.g. a recently abandoned well may be difficult to distinguish from a suspended or active well).

Data augmentation

We agree that further data augmentation experiments could be helpful. We will aim to evaluate several other simple techniques in the camera-ready version, in addition to resizing and flipping.

Accessibility of Planet data

Thank you for raising this excellent point. We agree that for certain users and application settings, Planet data will be less accessible than Landsat / Sentinel imagery. However, a large number of research institutions already have subscriptions to Planet imagery. Furthermore, for many use cases, a user may wish to target a relatively narrow area (e.g. abandoned wells within a regional jurisdiction that can be targeted for plugging). For some stakeholders, it also seems likely that applying algorithms to very large areas will be limited by computational constraints (given the relatively high resolution of the images), independent of data access.

We very much appreciate your suggestion to use Planet data to train teacher models, then derive student models for lower resolution publicly available data. This sounds like a fruitful follow-up paper, which we would be happy to mention in the conclusion.

Failure analysis:

Thank you for raising this point. In the camera-ready version, we will include representative failure cases from visually complex scenarios. Most patches contain only 1–5 wells, and we observe performance degradation in rare, high-density regions. Abandoned and suspended wells—due to their subtle visual signatures and lack of surface infrastructure—can be especially challenging to detect. Enhancing model performance in these edge cases is a priority for future work.

Thank you for your thoughtful and constructive review. We're grateful for your recognition of the dataset's potential impact on climate change mitigation, as well as for your helpful suggestions. Below are our responses to the points you raised:

Transformer-based models for well detection

Thank you for bringing this up. In our initial detection benchmarks, we focused on lightweight models (like SSD Lite and FCOS) because of their practical deployment relevance. However, we agree it is valuable to include transformer-based detection models for a more comprehensive comparison. We have performed experiments with the transformer-based model DETR, using a ResNet50 backbone. As shown below, DETR achieves strong localization performance, further validating the relevance of transformer-based models for this task.

Table R1 : Object detection results on the test set. We report Intersection over Union (IoU) at thresholds 0.1, 0.3, and 0.5, as well as mean Average Precision (mAP) at IoU = 0.5 and IoU ∈ [0.5, 0.95].

We also appreciate the reviewer’s point regarding the statement on Page 7, Lines 361–363 comparing object detection and segmentation performance. While earlier detection models such as RetinaNet and Faster R-CNN underperformed compared to segmentation counterparts, our updated DETR results demonstrate that transformer-based detectors can help close this gap. We will revise the manuscript to present this comparison more carefully and to avoid overgeneralizing across architectures.

Clustering algorithm

The goal of our clustering algorithm is to allow for a dataset split that avoids spatial autocorrelation across splits while still including many different locations in each split. By construction, our clusters demonstrate spatial coherence, and as shown in Figure 2 they are geographically distributed. (Note that since the wells are unevenly distributed across Alberta, a simple gridding approach would lead to significant imbalances between splits.) While we would be interested in analyzing our algorithm’s efficacy across similar geospatial datasets, we feel this would distract from the principal focus of the present paper and is not necessary to establish that the dataset splits are reasonable.

Annotation granularity and quality control

Thank you for raising this important point. We use standardized 90m circular masks and square bounding boxes for several reasons. Firstly, well pads are relatively standardized in size and shape; the 90m circular shape is typical, not merely a placeholder. Secondly, the very large size of our dataset would make the creation of human-annotated masks challenging, especially given the expert knowledge required for annotation in many cases. Finally, the imagery we use includes a near-infrared channel (and certain well pads are not visible in RGB alone), which makes human interpretation of the images harder.

Despite the coarse annotations, our results show that modern models achieve strong performance. This aligns with prior work (e.g., Rolnick et al., 2017) showing that deep learning models are robust to modest label noise. We will be exploring partial human annotations in future versions of the dataset, in collaboration with domain experts.

Figure readability

Thank you for the suggestion. We will revise Figure 3 and other figures by reorganizing subpanels, standardizing annotations, and increasing resolution to better highlight qualitative comparisons between the ground truth and predictions.

Use of metadata for multimodal extensions

Thank you for the suggestion. However, due to licensing, we can release imagery but not metadata. We’re actively exploring similar tasks using open data sources (e.g., Sentinel) to support future multimodal benchmarks.