Algorithm to Identify Frac Pits and Well Pads Using Satellite Map Data

Project Description

BACKGROUND

We are an online marketplace and data source for oilfield water data. In the past year we have started to analyze satellite imagery for oilfield water related features to include these results in our GIS map and marketplace products. We currently use an outside service company to identify these features. We would like to hire a contractor to build our own automated "virtual machine" to perform this analysis so that we can reduce operating costs of the satellite scans and increase their frequency and capabilities without relying on an outside service provider in the future, and also to own our IP from our investment in advancing these methods.

The key region we are covering is about 75,000 sqmi or 200km2. 

We are identifying several types of features.

FEATURE 1: WELL PADS

Well pads are square or rectangular areas of cleared, leveled dirt where drilling and fracturing equipment can be placed to drill a new oil well. Samples shown below. Dimensions may be 50m2 up to 100m2 or more. When they are newly constructed they have colace spread on them which increases reflectivity. A new dirt road must be built to reach a pad area before the pad can be cleared. We have had very high sensitivity levels using CNN-ML for detection (probably 85-90% detection rate) but also high false positive rates (probably 3:1 FP:TP ratio) — but this was just a first pass with no additional tuning or OBIA rules.

We need to:

a) Identify presence of a well pad. We are using 10m resolution Sentinel-2 imagery.

b) Distinguish well pad from non well pad similar features like a building construction site, road intersection, or harvested farm. We might be able to apply some OBIA rules to improve results, for example well pads have a dirt road leading to them and tend to be near other oil wells.

c) We do not care about well pad size or other qualities except to the extent these help our detection and specificity.

d) Well pad detection most occur frequently and generate results quickly — this is because the sole value of well pad detection is in identifying new activity on the ground and this new activity becomes widely known and without value in a few weeks.

Samples: https://screencast.com/t/g5g5bdrqc  and  https://screencast.com/t/JM2DXnorZ

FEATURE 2: FRAC PITS aka Frac Ponds aka Water Impoundments. 

These are artificial ponds used to supply hydraulic fracturing activities. Sample images below. Methods for identifying water in satellite imagery using Near I/R reflectivity are well-established. We need to:

a) Identify presence of a frac pit  (currently we use OBIA methods for this but ML could work as well or better). We are using 10m resolution Sentinel-2 imagery.

b) Distinguish frac pit from non frac pit water features such as: reserve pits (small water pits used for drilling operations, which should be classified separately), lakes and ponds, swimming pools, reservoirs, streams, wastewater treatment ponds etc. The identification and differentiation of water features can be done through a combination of ML and OBIA rule sets. For example, a frac pit that is nowhere near any known oil well locations might not be a frac pit, or a frac pit in the middle of a residential area might not be a frac pit.

b) Estimate its dimensions (typical size range is 75m2 - 200m2)

c) Estimate volume either based on a standard pit design methods (simple conversion from surface area on a formula) — if we can find a more advanced way to estimate volume that would be even better.

d) Estimate water type: fresh, brackish or produced water (we have found this can be done through turbidity or color or spectral analysis but have not tested reliability)

e) Track changes in same-pit levels over time.

f) Ideally, identify early signs of frac pit construction, not just presence of water.

Sample frac pit images: https://screencast.com/t/BHgBh7ZBqQti

FUTURE WORK: We have a number of other features we would like to detect in addition to frac pits and well pads. Almost everything that happens on the ground in the Permian Basin is of interest to us. If we hire you to do this work for us we will want to add many more feature categories and we have several in mind already.

SYSTEM CAPABILITIES

IMPORTANT: Well Pad identification is the more urgent requirement at this time. Well pad process is priority 1. Frac pits is priority 2. The well pad detection system below should be built as expeditiously as possible and can operate independently of the frac pit detection system. If they go together, great, but frac pit development should not delay or add initial cost to well pad development.

1) Automatically download, combine and load Sentinel-2 (free, 10m resolution) satellite imagery for the regions of interest on whatever frequency these images become available. I believe this frequency to be about once per week, however I am not sure the entire region of interest is covered with every pass.

• If necessary for early versions this can be done by a human operator and not automated.
• In the future we might upgrade our image resolution and frequency to more expensive sources. This might prove necessary to identify some of the features we want to identify in the future.

2) Scan the imagery for the desired features described above using ML and OBIA models to be developed and improved over time.

3) Output the GPS coordinate locations of all well pads and frac pits (include dimensions and estimated volume and water type with frac pits) and other features when added

4) Compare newest scan to all prior scans to screen out all previously identified features and highlight only the new features that have appeared in the most recent scan period

5) Construct and operate a system for human QA of new features. Something like a Captcha system for image confirmation and classification so that all newly identified features are shown in series to trained human QA persons in a remote location for confirmation, rejection, enhancement or reclassification

6) A method for feeding the results of human QA back into the ML model and OBIA rules to improve them, and also output the post-QA feature locations list as the final list for this scan date in an appropriate database format.

7) Match the resulting locations to various external data sets to enhance their meaning and for further QA. For example, we should match newly identified well pads to the most recent state data on drilling permit locations in order to remove well pad locations for which permits have already been filed, since these pads can already be found by other means. Also matching new pad locations to the closest published operating drilling permits in order to estimate whi company is the likely owner of the new pads.

YOUR PROPOSAL TO BUILD A MINIMUM VIABLE PRODUCT

In your proposal, please answer the following four questions:

1) What are your background and qualifications for this project?

2) What would be your approach to solving this problem?  Which algorithms are appropriate?

3) Will you work be limited to the development of the algorithm or can you develop and deploy a full solution in the cloud?  We are more concerned about performance than aestethics of the user interface.

4) How would you estimate the cost of Feature 1 and Feature 2?  Can you provide a flat fee quote for each?