Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Scientific Data volume 11, Article number: 67 (2024 ) Cite this article Cracking Solution
Remotely sensed imagery has increased dramatically in quantity and public availability. However, automated, large-scale analysis of such imagery is hindered by a lack of the annotations necessary to train and test machine learning algorithms. In this study, we address this shortcoming with respect to above-ground storage tanks (ASTs) that are used in a wide variety of industries. We annotated available high-resolution, remotely sensed imagery to develop an original, publicly available multi-class dataset of ASTs. This dataset includes geospatial coordinates, border vertices, diameters, and orthorectified imagery for over 130,000 ASTs from five labeled classes (external floating roof tanks, closed roof tanks, spherical pressure tanks, sedimentation tanks, and water towers) across the contiguous United States. This dataset can be used directly or to train machine learning algorithms for large-scale risk and hazard assessment, production and capacity estimation, and infrastructure evaluation.
John Francis, Jonathan Bright, … Vincent J. Straub
Dew Nahime Towers & Piero Fraternali
L. Kruitwagen, K. T. Story, … C. Hepburn
Many vital industries, including petroleum and chemical production, processing, refining, and transport, rely on short-term storage in above-ground storage tanks (ASTs)1. For example, in the United States, the total petroleum storage capacity at sites that commonly use ASTs exceeds two billion barrels2. Typical types of ASTs include closed roof, external floating roof, and spherical pressure tanks, the choice of which depends on the characteristics of the stored material, including the specific gravity, volatility, and flash point3,4.
ASTs are vulnerable to various hazards, both natural and anthropogenic5. For example, during Hurricanes Katrina6, Isaac7, and Harvey, thousands of barrels of petroleum were released from storage tanks and spread by floodwaters into the environment8,9,10, threatening ecosystems, economies, and human health11,12,13. Internationally, oil depots have been targeted during civil unrest and military conflict, demonstrating the possibility for ASTs to become targets14,15. In 2020, as a result of the covid-19 pandemic and the Russia-Saudi Arabia price war, the crude oil market experienced decreased demand and increased inventory, straining storage capacity16. Despite these widespread risks, consolidated data suitable for assessing system vulnerabilities and failures, estimating current production and capacity, and evaluating the state of energy and other infrastructure are not readily available to researchers, regulators, and other decision-makers.
The Environmental Protection Agency (EPA) collects information on facilities that process petroleum and hazardous materials through the Facility Registry System (FRS) and Toxic Release Inventory (TRI); however, these datasets do not contain specific tank locations, sizes, or contents17. The Emergency Planning and Community Right-to-Know Act (EPCRA) requires that industrial sites make chemical inventory data available to community members via material safety data sheets (MSDS)18; however, these data are not available outside of one’s immediate locale.
Remotely sensed imagery has been utilized in some cases to develop AST datasets for specialized purposes. The Oil and Gas Storage Tank (OGST) dataset, for example, identifies tanks located inside the footprint of well pads across Alberta, Canada, in 760 images with resolutions varying from 0.3 to 1.2 meters19. The Oil Storage Tank (OST) dataset of external floating roof tanks was developed to estimate capacity and contains 1,595 annotated images taken from Google Earth20. A dataset of over 4,500 storage tanks along the Houston Ship Channel (HSC), includes tank location and indication of contaminant barriers, as well as estimates of diameter, height, shell thickness, and content characteristics21. Finally, the DOTA dataset includes over 10,000 annotated storage tanks from aerial imagery at various resolutions and locations worldwide22. While these datasets all provide useful data in certain contexts, they are generally limited by too few annotations, missing location data, lack of geographic coverage in the United States, simplified classifications, and limited availability.
Recognizing the value of comprehensive, publicly available data for community health and safety risk assessment, petroleum market research, and other analyses, we have developed an original dataset of ASTs from high-resolution aerial imagery across the contiguous United States. It contains geospatial coordinates, border vertices, and orthorectified imagery for over 130,000 spherical pressure, closed roof, external floating roof, sedimentation tanks, and water towers. This dataset was developed for two primary purposes: (1) as training and testing data for subsequent object detection algorithms, and (2) as geospatial data for AST risk assessments. Additionally, the dataset will benefit researchers interested in:
Petroleum storage working and net capacity estimates;
Petrochemical market evaluation and economic assessment; and
Machine learning or computer vision tasks, particularly those that require extensive training datasets.
We constructed this dataset to contain the imagery, location, bounding boxes, and classifications for various ASTs across the contiguous United States. As illustrated in Fig. 1, our process involved: (1) selecting aerial imagery data, (2) manually annotating and classifying storage tanks, water towers, and sedimentation tanks, (3) validating and correcting when necessary, the bounding boxes and classifications, (4) obtaining geospatial information for each tank, image, and tile, and (5) developing tile level annotations and compiling the full dataset. Each step will be described in the subsections that follow.
Schematic of the data development and validation process.
The United States Department of Agriculture (USDA) Farm Service Agency’s (FSA) National Agriculture Imagery Program (NAIP) collects high-resolution, remotely-sensed aerial imagery of the continental U.S. during the agricultural growing seasons23. We selected NAIP tiles based on two criteria: (1) presence of relevant infrastructure and (2) objects being clearly identifiable. The first criterion was evaluated using existing point datasets of natural gas and petroleum processing plants, reserves, terminals, and ports from the U.S. Department of Homeland Security, Energy Information Administration (EIA), and National Oceanic and Atmospheric Administration (NOAA)24,25,26,27,28,29,30. As objects of interest range from 3 to 69 meters in diameter, to meet the second criterion only NAIP tiles collected after 2018 were included, guaranteeing a minimum 60 cm ground sampling distance (GSD) so that objects are represented by multiple pixels23,31.
A total of 2132 NAIP tiles from 48 states were acquired for annotation from the Microsoft Planetary Computer Data Catalog repository32. Each NAIP tile includes four bands (red, green, blue, and near-infrared) over a 3.75-minute longitude by 3.75-minute latitude quarter quadrangle with an additional 300-meter buffer, thereby covering between 12 and 18 square miles23. The large size, high-resolution, and high concentration of objects require meticulous annotation33. To ensure adequate visual inspection, each NAIP tile was broken into smaller 512-by-512-pixel images, representing approximately 23 acres (0.094 square kilometers) (Fig. 2).
Example of tile (m_4107341_sw_18_060_20190917) split into 512 by 512-pixel images with row and column indices.
File names for tiles utilize the USDA’s original file naming convention separating the USGS quadrangle identifier corresponding to a 7.5-minute × 7.5-minute area; quarter-quad two-character ordinal direction identifying a given quarter section within a quadrangle; two-digit Universal Transverse Mercator (UTM) zone; image resolution; and capture date in year-month-day, each separated by underscores. 512-by-512-pixel images and image-level annotations utilize the same naming convention and include the row and column index corresponding to the position of the image within the tile (Fig. 3).
Interpretation of NAIP tile and image file naming convention.
To accurately and efficiently review and annotate the large number of images identified, we adapted the large-scale annotation procedure developed to create ImageNet34,35. Using the NAIP imagery, research assistants manually annotated images using LabelImg, an annotation tool with a graphical user interface (GUI) written in Python36. LabelImg allows the researcher to view and identify positive images, or images containing objects of interest, and manually draw rectangular bounding boxes. Annotations and images were further inspected using the zoom in/out function, adjusted using a sidebar with annotation details, or deleted if created in error.
Research assistants annotated tanks into one of five categories (Fig. 4). External floating roof tanks are cylindrical shells outfitted with a roof that floats on the surface of the stored liquid, rising and falling with the liquid level, and are suitable for liquids with high vapor pressure. By contrast, closed-roof tanks have coned, domed, or flat roofs permanently affixed to the tank shell, and generally contain products with lower vapor pressures. Closed roof tanks may also have an internal floating roof. To allow versatility for end users, closed roof tanks less than or equal to 9 meters in diameter are classified as narrow closed roof tanks. Spherical pressure tanks, also known as Horton spheres, are designed to contain gases or liquids at pressures significantly different than the external environment and to withstand significant internal pressures. Although of less interest for our purposes, water towers and sedimentation tanks were annotated due to their visual similarity, and in some cases proximity, to ASTs of primary interest. Water towers are tanks elevated to provide pressure for water distribution, while sedimentation tanks are generally used in water and wastewater treatment.
Data samples of each tank type category in the dataset.
To maintain consistency, research assistants were provided with training on the visual differences between object classes, the LabelImg software, and image annotation protocol. To prevent worker fatigue, research assistants were allocated subsets of the images to annotate on a weekly basis. Due to the low ratio of positive-to-negative images, research assistants reviewed only the positive images and annotations in each subset. The annotation and validation required 22 research assistants to review 22,815 images over 2,300 hours.
Using the orthorectified NAIP tiles, longitude and latitude coordinates were identified for the northwest and southeast vertices for images and objects. A mesh overlay of the Universal Transverse Mercator (UTM) coordinates was linearly interpolated using the northwest and southeast tile vertices and then transformed into longitude and latitudes using the tile datum. The tile pixel coordinates were then used to obtain each object’s northwest and southwest longitude and latitude coordinates.
Annotated images themselves provide valuable data for a variety of contexts; however, some analyses, particularly infrastructure risk assessment, require data compiled at larger spatial scales and the estimation of additional characteristics, such as location and diameter. Tile level annotations were created to provide this information by translating the image’s pixel coordinates, as well as the row and column position within the tile, to tile pixel coordinates37. Approximately 20 percent of the objects in the image level annotations are truncated objects or objects that partially lie outside of the image. Within each tile, truncated objects were merged if they belonged to the same class, separated by a maximum of three meters, and did not intersect. The diameter of each object was calculated as the minimum width or height of each object’s bounding box. Additionally, the capture date is included to allow for longitudinal analyses when multiple tank annotations may exist over time.
The individual 512 by 512-pixel images and corresponding annotations were compiled into a nationwide dataset along with metadata37,38. The annotated dataset includes 142,107 objects distributed across seven classes. Due to the image sampling methods or natural object frequencies, the validated dataset is imbalanced in favour of closed roof and narrow closed roof tanks, which comprise 51% and 35% of the dataset, respectively (Table 1).
Figure 5 depicts the distribution of annotated objects across the contiguous United States in comparison to active terminals and refineries reported to the Internal Revenue Service39,40. This figure illustrates that the spatial coverage of the above-ground storage tank dataset is generally consistent with existing petroleum datasets that were not used to develop our dataset.
Fishnet grid of the density of annotated tanks in comparison to CONUS and IRS reported active terminals and refineries.
Separate data products, including the aerial imagery dataset38, above-ground storage tank inventory37, and image and tile characteristics41,42, totaling 2.3 GB in size, are available in a Figshare repository43. Each component is detailed below.
The above-ground storage tank aerial imagery dataset provides a ready-to-use training dataset for deep learning models. This dataset is comprised of two folders which contain 27639 512-by-512-pixel images from 2132 tiles stored in jpg format (“images” folder) and the corresponding annotations in PASCAL Visual Object Classes (VOC) 2007 format as Extensible Markup Language (XML) files (“xmls” folder).
The above-ground storage tank inventory dataset provides cleaned geospatial data and storage tank characteristics for tanks across CONUS. The table is accessible in JSON, GeoJSON, and ESRI shapefile formats to allow for accessibility across applications and software. This dataset is accompanied by a description of the variables included in the inventory dataset.
Metadata relating to the images and tiles are provided by two data records accessible in CSV file format. The image characteristics metadata provides the name of each image and the N.W. and S.E. latitude and longitude coordinates. Additionally, each image can be related to the corresponding parent NAIP tile using the tile name, the image’s row and column index within the tile, and the N.W. and S.E. pixel coordinates respective to the tile. The tile characteristics data record provides the tile name, latitude, and longitude extent of the tile. The image and tile characteristics data records are supplemented with a description of their contents, including the field name, description, data type and units.
The development of a high-value, large-scale, multi-class dataset through manual annotation can be laborious and expensive. To efficiently and cost-effectively ensure data quality, we implemented a validation procedure originally developed and described in further detail by Su, 201234. In this validation procedure, we controlled for three potential sources of error: (1) missed objects; (2) poorly drawn bounding boxes; and (3) misclassified objects.
Missed objects. Relying on human annotators requires that annotators label every object in every image. This is made difficult by potential changes in the appearance of objects due to imagery characteristics such as sensor aperture, cloud characteristics, and time of day when collected.
Poorly drawn bounding boxes. Rectangular bounding boxes provide the spatial location for an object of interest in an image. When a bounding box is too large, excess pixels are incorrectly noted as part of an object, most commonly occurring with smaller objects. Alternatively, when a bounding box is too small, pixels representing an object of interest are incorrectly excluded (i.e., the legs of water towers or the walls of tanks).
Misclassified objects. Annotators may confuse: (1) objects of interest for one another, such as external floating roof tanks and sedimentation tanks, or (2) objects of interest with other objects in the image, such as circular buildings, concrete pads, or above-ground pools. In addition, annotators may neglect to indicate when objects are truncated by the edge of the image.
Several procedures were implemented to minimize the effect of these errors and ensure that quality was maintained across all classes and annotators.
Sorokin and Forsyth note that worker compensation may affect the quality of the data annotations, where low wages workers may lose motivation, and high wages would waste resources and attract inefficient workers44. Therefore, we controlled the quality of our data set by hiring student research assistants (RAs), allowing our team to carefully screen applicants and employ highly skilled workers who are jointly incentivized by wages and obtained experience.
RAs were given extensive training to learn the purpose of the project, use LabelImg, properly annotate images, and understand and distinguish between the object classes. During the training, RAs were instructed to label every object of interest, including those truncated by the edge of an image. To prevent false negatives or mislabeled annotations, annotators were instructed to label ambiguous objects of interest as undefined objects so they may be further reviewed during validation. After the completion of the first subset of images and periodically during the annotation phase, RAs were provided with additional training and feedback on their annotations and opportunities to make corrections to minimize the propagation of systematic errors. Additionally, RAs had access to a Slack channel where they could ask for clarification on the annotation procedure, receive feedback on a specific image, and learn from guidance provided to other RAs.
All images and corresponding annotations were reviewed by a set of three RAs using a simple workflow illustrated in Fig. 1. The first RA ensured that every object instance in an image has a corresponding bounding box. This ensures that objects in an image that may have been missed are annotated, particularly small objects, images truncated by the edge of an image, or those outside a cluster of objects. Additionally, they removed bounding boxes drawn in error. The next RA reviewed the quality of each bounding box and adjusted it where necessary to ensure the tightest possible bounding box around every object in each image. The final RA confirmed each annotation was correctly classified, reclassified undefined objects if possible, and checked that truncated objects were correctly marked.
Since consistency across annotators is a positive characteristic of a dataset, we used multiple metrics to evaluate the changes resulting from the validation procedure. To assess the coverage validation, the number of bounding boxes added and removed during validation was calculated (as illustrated in Fig. 6). The pre- and post-validation datasets contain 128,127 and 142,091 objects, respectively, and it was determined that during the coverage validation process, 0.8 percent (1,067 objects total) of the pre-validation and 10.6 percent (15,032) of the post-validation annotations were removed and added, respectively, as shown in Table 2. To evaluate the quality validation procedure, the number of pre- and post-validation bounding box agreements and exact matches were determined. Of the objects in the post-validation dataset, 89.4 percent (127,050) of the bounding boxes were in agreement and 78.6 percent (111,694) matched exactly with the pre-validation boxes (Table 2). Thus, the difference between bounding boxes in agreement and that exactly match, 10.8 percent (15,356) of the post-validation bounding boxes, required minor adjustments.
Illustration of the coverage and quality assessment criteria. On the left, pre-validation annotations are shown on top and the post-validation annotations are shown below. Bounding boxes were determined to be in agreement through a multi-step process. For each post-validation bounding box, the intersection-over-union (IoU) was calculated with respect to each of the pre-validation bounding boxes. The highest IoU for a post- and pre-validation bounding box pair was identified. (a.1) If the IoU exceeded or equalled 0.5, the post- and pre- validation bounding box pair was determined to be in agreement. (a.2) Furthermore, if the IoU was 1.0, the bounding box pair was regarded as an exact match. (b) If the highest IoU for a bounding box was less than 0.5 and the bounding box was present only in the pre-validation or post-validation annotations, we determined that it was removed or added, respectively, during the validation process.
In analysing the bounding box agreement of the annotations, we did not consider the class labels (tank types). Looking at both the agreement of bounding boxes and the assignment of class labels, 93.8 percent (133,303) are in agreement with the pre-validation bounding boxes and have the same class label, while 73.4 percent (104,344) of bounding boxes exactly match and have the same class with respect to the post-validation bounding boxes. To assess the number of changes made to class labels during the class validation phase, the joint frequency of the pre- and post-validation classes for bounding boxes in agreement were tabulated as shown in Table 3. The values on the diagonal of the table indicate the percent of class labels matching pre- and post-validation. The sum of the diagonals, 93.1 percent of the bounding boxes in agreement, indicates that the vast majority of class labels matched before and after validation. The largest share of those that changed class were between closed roof tanks and narrow closed roof tanks. Seven percent of the tanks believed to be closed roof pre-validation were converted to narrowed closed roof tanks, and 5.2 percent were changed vice versa. Given the similarity between these two classes, these changes are not surprising.
Further evaluation of the efficacy of the large-scale validation procedure was conducted on a subset of the data. One percent of the compiled dataset, or 276 images, were randomly selected to be judged according to what we are calling a ground truth dataset. The lead author, with expertise in AST risk assessment, corrected and reviewed each image. To ensure that objects in each image were correctly attributed to the right class, the expert utilized their knowledge, Google Maps, and Google Street View to locate and classify each tank as accurately as possible. This process of checking multiple sources for each tank was far more time-consuming and so is not scalable. However, it provided another assay of the quality of the annotation and validation procedure.
Results from comparing the ground truth data to the results of the validation process revealed an average precision of 0.99 and recall of 0.952 across tank classes. These indicate that the bounding boxes were correctly associated with objects of interest, and fewer than 5 percent of objects were missed. Further inspection offered insight into the causes of missed objects. Of the bounding boxes that were missed, 71.9 percent had a pixel area less than 162 (i.e., an area less than 92.2 square meters) and 93 percent had a pixel area less than 322 (i.e., an area less than 368.64 square meters), indicating that smaller tanks were harder to identify.
The raw aerial imagery and annotation tools used in this study are publicly accessible36,45. The source code developed by the authors to process the imagery and develop the tank inventory dataset are available on GitHub (https://github.com/celinerobi/ast-data-pipeline).
Hildebrand, M. S., Noll, G. G. & Hand, W. T. Above ground bulk storage tank emergencies.
Working and Net Available Shell Storage Capacity - U.S. Energy Information Administration (EIA). https://www.eia.gov/petroleum/storagecapacity/.
Myers, P. E. Aboveground storage tanks. (McGraw-Hill Education, 1997).
Chattopadhyay, S. Pressure Vessels: Design and Practice. (CRC Press).
U.S Department of Energy. Effect of Sea Level Rise on Energy Infrastructure in Four Major Metropolitan Areas. 44 (2014).
United States Government Accountability Office. Hurricane Katrina: EPA’s Current and Future Environmental Protection Efforts Could Be Enhanced by Addressing Issues and Challenges Faced on the Gulf Coast. www.gao.gov/cgi-bin/getrpt?GAO-07-651 (2007).
Lessons From Hurricane Isaac: Gulf Coast Coal and Petrochemical Facilities Still Not Storm Ready. Gulf Monit. Consort. Rep.
Campedel, M. Analysis of Major Industrial Accidents Triggered by Natural Events Reported in the Principal Available Chemical Accident Databases. JRC Sci. Tech. Rep. (2008).
Myers, P., PEMY Consulting LLC & American Petroluem Institute. API 656 Storage Tank NATECH Natech(Natural Hazard Triggered Technological Accidents). (2021).
Union of Concerned Scientists & Texas Environmental Justice Advocacy Services. Double Jeopardy in Houston: Acute and Chronic Chemical Exposures Pose Disproportionate Risks for Marginalized Communities. 1–25 (2016).
Girgin, S., Necci, A. & Krausmann, E. Natech hazard and risk assessment. in Words into Action Guidelines: National Disaster Risk Assessment (United Nations Office for Disaster Risk Reduction, 2017).
Nascimento, K. R. D. S. & Alencar, M. H. Management of risks in natural disasters: A systematic review of the literature on NATECH events. J. Loss Prev. Process Ind. 44, 347–359 (2016).
Cozzani, V. et al. Quantitative assessment of domino and NaTech scenarios in complex industrial areas. J. Loss Prev. Process Ind. 28, 10–22 (2014).
Khalifa Haftar. Forces loyal to Khalifa Haftar attack ports in Libya. Al Jazeera (2016).
Hassan, A., Pugmire, J. & Gambrell, J. Yemen rebels strike oil depot in Saudi city hosting F1 race | AP News. Associated Press (2022).
Ma, R. R., Xiong, T. & Bao, Y. The Russia-Saudi Arabia oil price war during the COVID-19 pandemic. Energy Econ. 102, 105517 (2021).
Article PubMed PubMed Central Google Scholar
Santella, N., Steinberg, L. J. & Sengul, H. Petroleum and hazardous material releases from industrial facilities associated with hurricane katrina. Risk Anal. 30, 635–649 (2010).
Title 42, Chapter 116 - Emergency Planning and Community Right-To-Know. vol. 42. U.S.C 1–26 (2015).
Rabbi, J., Ray, N., Schubert, M., Chowdhury, S. & Chao, D. Small-object detection in remote sensing images with end-to-end edge-enhanced GAN and object detector network. Remote Sens. 12, 1–25 (2020).
Heyer, K. Oil Storage Tanks. Kaggle https://www.kaggle.com/datasets/towardsentropy/oil-storage-tanks (2019).
Bernier, C., Elliott, J. R., Padgett, J. E., Kellerman, F. & Bedient, P. B. Evolution of Social Vulnerability and Risks of Chemical Spills during Storm Surge along the Houston Ship Channel. Nat. Hazards Rev. 18, 04017013 (2017).
Xia, G.-S. et al. DOTA: A Large-scale Dataset for Object Detection in Aerial Images *. (2019).
Earth Resources Observation And Science (EROS) Center. National Agriculture Imagery Program (NAIP)., https://doi.org/10.5066/F7QN651G (2017).
Energy Information Administration (EIA). U.S. Energy Atlas - Natural Gas Processing Plants. https://atlas.eia.gov/datasets/natural-gas-processing-plants/.
Energy Information Administration (EIA). U.S. Energy Atlas - Petroleum Refineries. https://atlas.eia.gov/datasets/petroleum-refineries/.
Energy Information Administration (EIA). U.S. Energy Atlas - Northeast Petroleum Reserves. https://atlas.eia.gov/datasets/northeast-petroleum-reserves.
Energy Information Administration (EIA). U.S. Energy Atlas - Petroleum Product Terminals. https://atlas.eia.gov/datasets/petroleum-product-terminals/.
Energy Information Administration (EIA). U.S. Energy Atlas - Strategic Petroleum Reserves. https://atlas.eia.gov/datasets/strategic-petroleum-reserves/.
United States Department of Homeland Security. Petroleum Ports for the Homeland Infrastructure Foundation-Level (HIFLD) Database - Petroleum Terminals. https://hifld-geoplatform.opendata.arcgis.com/datasets/petroleum-ports/.
National Oceanic and Atmospheric Administration (NOAA). Major U.S. Port Cities and Satellite Ports FY 1999/2000. https://www.ngs.noaa.gov/RSD/coastal/projects/coastal/ports_list.html.
United States Department of Agriculture. NAIP Coverage 2002–2018. (2019).
Planetary Computer Data Catalog. NAIP: National Agriculture Imagery Program.
Torney, C. J. et al. A comparison of deep learning and citizen science techniques for counting wildlife in aerial survey images, https://doi.org/10.1111/2041-210X.13165 (2019).
Su, H., Deng, J. & Fei-Fei, L. Crowdsourcing annotations for visual object detection. AAAI Workshop - Tech. Rep. WS-12-08, 40–46 (2012).
Deng, J. et al. ImageNet: A large-scale hierarchical image database. in 2009 IEEE Conference on Computer Vision and Pattern Recognition 248–255. https://doi.org/10.1109/CVPR.2009.5206848 (2009).
Robinson, C., Bradbury, K. & Borsuk, M. E. Above Ground Storage Tank Inventory Dataset. Figshare https://doi.org/10.6084/m9.figshare.22665787 (2023).
Robinson, C., Bradbury, K. & Borsuk, M. E. Above Ground Storage Tank Aerial Imagery Dataset. Figshare https://doi.org/10.6084/m9.figshare.22543456 (2023).
Internal Revenue Service. Refinery Control Number RCN Refinery Location Directory.
Internal Revenue Service. Terminal Control Number TCN Terminal Locations Directory.
Robinson, C., Bradbury, K. & Borsuk, M. E. Image Characteristics. Figshare https://doi.org/10.6084/m9.figshare.22661005 (2023).
Robinson, C, Bradbury, K. & Borsuk, M. E. Tile Characteristics. Figshare https://doi.org/10.6084/m9.figshare.22661011 (2023).
Robinson, C, Bradbury, K. & Borsuk, M. E. Remotely sensed above-ground storage tank dataset for object detection and infrastructure assessment. Figshare https://doi.org/10.6084/m9.figshare.c.6649661 (2023).
Sorokin, A. & Forsyth, D. Utility data annotation with Amazon Mechanical Turk. 2008 IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. Workshop CVPR Workshop 1–8, https://doi.org/10.1109/CVPRW.2008.4562953 (2008).
Planetary Computer Hub. Accessing NAIP data with the Planetary Computer STAC API.
Student research assistants performed the manual annotation of storage tanks, sedimentation tanks, and water towers using the LabelImg GUI. In particular, we thank the diligent efforts of Qianyu Zhao, Sunny Li, Jackson Harwood, Joshua Kang, Jaewon Jong, Dominic Van Cleave-Schottland, Ryan Feinberg, and Chunxin Tang who served for months as manual annotators and/or validators for this work. This work was supported in part by the Alfred P. Sloan Foundation grant number G-2020-13922 and in part by Environmental Protection Agency (EPA) STAR grant number R840041. The content is solely the responsibility of the authors and does not necessarily represent the official views of the Alfred P. Sloan Foundation or the EPA.
Department of Civil and Environmental Engineering and Duke Center on Risk, Duke University, Durham, North Carolina, 27708, USA
Celine Robinson & Mark E. Borsuk
Department of Electrical and Computer Engineering and Nicholas Institute for Energy, Environment & Sustainability, Duke University, Durham, North Carolina, 27708, USA
You can also search for this author in PubMed Google Scholar
You can also search for this author in PubMed Google Scholar
You can also search for this author in PubMed Google Scholar
C.R. led this research, created the original idea for the AST detection project, developed the algorithm to merge polygon annotations, developed validation procedure, led the student research team management from Summer 2021 to Spring 2022, was involved in all stages of data collection, assembly, and management, and wrote most of the manuscript. M.B. provided intellectual guidance and edited the manuscript. K.B. provided intellectual leadership on the project, georeferencing, and edited the manuscript.
The authors declare no competing interests.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Robinson, C., Bradbury, K. & Borsuk, M.E. Remotely sensed above-ground storage tank dataset for object detection and infrastructure assessment. Sci Data 11, 67 (2024). https://doi.org/10.1038/s41597-023-02780-1
DOI: https://doi.org/10.1038/s41597-023-02780-1
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.
Provided by the Springer Nature SharedIt content-sharing initiative
Scientific Data (Sci Data) ISSN 2052-4463 (online)
Cracking And Depolymerization Reaction Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.