Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation...

8
Orbital Insight Energy: Oil Storage v5.1 Methodologies & Data Documentation Overview and Summary Orbital Insight Global Oil Storage leverages commercial satellite imagery, proprietary computer vision algorithms, and advanced data science methodology to provide daily crude oil estimates at global, regional and country-specific levels. This document describes the practice and methodology we use to produce the energy dataset. The Global Oil Storage dataset presents a time series of full-volume estimates based on billions of barrels of crude oil that is stored in tens of thousands of floating-roof tanks (FRTs) around the world on a daily basis. We apply artificial intelligence to satellite imagery in order to locate and dimension each floating-roof tank, generating a database of FRTs worldwide. Using the trigonometry of FRT shadows to estimate the fill volume of oil within each tank, we calculate the aggregate volume of global oil storage. The resulting time series reflects the volume of FRTs grouped within regions of interest. Our dataset accounts for oil volume specifically stored in FRTs, excluding crude oil volume stored using other methods such as fixed-roof tanks, underground caverns, very large crude carriers (VLCCs), railcar, truck or pipeline storage. What problem does Orbital Insight Energy solve? Crude oil is one of the world’s most valuable commodities but is mired by a lack of transparency, often from opaque swing producers around the world — such as those within The Organization of Petroleum Exporting Countries (OPEC) and China. Storage estimates provided by organizations like the International Energy Agency (IEA) and Joint Organizations Data Initiative (JODI) are survey-based and result in delayed information (by weeks and often months) and credible reliability concerns. Orbital Insight uses a scientific approach rooted in satellite observation to bring visibility into oil storage regions across the globe. Changes in daily crude oil inventories are leading indicators of production and consumption over time, providing insight into the supply and demand forces that influence benchmark oil price changes. How does Orbital Insight Energy solve this problem? Crude Oil Storage Tank Identification & Dimensioning By leveraging information from public databases and our computer vision algorithms to locate tank farms and individual tanks, Orbital Insight has identified over 25,000 tanks worldwide — representing the world’s most comprehensive crude oil tank catalog. After we identify tank locations, a human worker manually labels tank images so that tank heights and diameters may be calculated. We use these dimensions to determine fill percentage as well as answer other questions regarding the maximum storage capacity of a tank farm, state, country, or geographic region. © Orbital Insight, Inc. v. 20181114

Transcript of Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation...

Page 1: Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation verview and Summary Orbital Insight Global Oil Storage leverages commercial satellite

Orbital Insight Energy: Oil Storage v5.1 Methodologies & Data Documentation

Overview and Summary Orbital Insight Global Oil Storage leverages commercial satellite imagery, proprietary computer vision algorithms, and advanced data science methodology to provide daily crude oil estimates at global, regional and country-specific levels. This document describes the practice and methodology we use to produce the energy dataset. The Global Oil Storage dataset presents a time series of full-volume estimates based on billions of barrels of crude oil that is stored in tens of thousands of floating-roof tanks (FRTs) around the world on a daily basis. We apply artificial intelligence to satellite imagery in order to locate and dimension each floating-roof tank, generating a database of FRTs worldwide. Using the trigonometry of FRT shadows to estimate the fill volume of oil within each tank, we calculate the aggregate volume of global oil storage. The resulting time series reflects the volume of FRTs grouped within regions of interest. Our dataset accounts for oil volume specifically stored in FRTs, excluding crude oil volume stored using other methods such as fixed-roof tanks, underground caverns, very large crude carriers (VLCCs), railcar, truck or pipeline storage. What problem does Orbital Insight Energy solve? Crude oil is one of the world’s most valuable commodities but is mired by a lack of transparency, often from opaque swing producers around the world — such as those within The Organization of Petroleum Exporting Countries (OPEC) and China. Storage estimates provided by organizations like the International Energy Agency (IEA) and Joint Organizations Data Initiative (JODI) are survey-based and result in delayed information (by weeks and often months) and credible reliability concerns. Orbital Insight uses a scientific approach rooted in satellite observation to bring visibility into oil storage regions across the globe. Changes in daily crude oil inventories are leading indicators of production and consumption over time, providing insight into the supply and demand forces that influence benchmark oil price changes.

How does Orbital Insight Energy solve this problem? Crude Oil Storage Tank Identification & Dimensioning By leveraging information from public databases and our computer vision algorithms to locate tank farms and individual tanks, Orbital Insight has identified over 25,000 tanks worldwide — representing the world’s most comprehensive crude oil tank catalog. After we identify tank locations, a human worker manually labels tank images so that tank heights and diameters may be calculated. We use these dimensions to determine fill percentage as well as answer other questions regarding the maximum storage capacity of a tank farm, state, country, or geographic region.

© Orbital Insight, Inc. v. 20181114

Page 2: Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation verview and Summary Orbital Insight Global Oil Storage leverages commercial satellite

Figure 1: Tank Farms in Cushing, Oklahoma with white circles marking identified FRTs (left); Individual tank metadata (right)

Finding Oil Tank Farms We have developed a tank farm finder algorithm to automate the task of locating tank farms in commercial satellite imagery. The tank farms cover more than 800 Area of Interest (AOI) polygons. Once a new tank is discovered, we manually mark the center of each tank (Figure 2) and store the associated polygon in our database.

Figure 2: Red dots denote the center or each FRT found within a tank farm

Human Labeling for FRT Images Orbital Insight has developed an internal image-labeling tool that we use to manually mark characteristics of the FRT in order to measure the FRT dimensions. A human worker fits circles around the observed tank structure and shadows to calculate the height and diameter of the FRT. Our proprietary trigonometric projections address variations in tank orientation and sun angles. The same image-labeling tool and shadow measurement methodology is also used to estimate the tank’s crude oil volume. Examples of the image-labeling tool are shown below in Figures 3 and 4.

© Orbital Insight, Inc. v. 20181114

Page 3: Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation verview and Summary Orbital Insight Global Oil Storage leverages commercial satellite

Tank Dimensions for Height Measurement

Figure 3: Workers label the floating roof tank by fitting circles to the bottom of the FRT and the

external shadow

Figure 4: Workers label the floating roof tank by fitting circles to the top and bottom of the tank as

well as the internal and external shadow

Volume Calculation By calculating the trigonometry of the circles fitted to the tank and tank shadows, we are able to determine the tank’s dimensions and fill volumes. Total volume of a cylinder-shaped tank (shell volume) = area of circular base X tank height Maximum storage capacity of tank (net shell volume) = shell volume —(roof thickness volume + base height) The filled volume of a tank is calculated in the same manner but uses the height of the floating roof instead of the shell height. We calculate the tank dimensions for all tanks and estimate the fill volume for a small proportional sample of tanks that will be used as training data for our computer vision algorithm.

Data Science Methodology Orbital Insight’s oil volume estimates are informed by the satellite imagery we ingest and process. We work with multiple commercial satellite imagery providers and obtain imagery in two ways:

1. We opportunistically sample provider catalogs and use every image collected that overlaps with any FRTs we track.

2. We task satellite providers and are assured imagery coverage for a given region and time.

The majority of imagery used to comprise our Global Oil dataset comes from opportunistic sampling; for any given day, we are typically able to observe up to 15% of the FRT population.

© Orbital Insight, Inc. v. 20181114

Page 4: Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation verview and Summary Orbital Insight Global Oil Storage leverages commercial satellite

The number of observations available continues to grow as additional satellites launch and become operational. For the remaining percentage of tanks not observed on a given day, we use advanced data science methods to estimate changes in fill percent. Our research has shown that the population of tanks for a given geographic region tends to be heterogeneous in terms of tank size, tank fill percent and volatility. Of course, the more tanks of the population we can observe, the better our estimate of the entire population can be. At the individual tank level, we build a model of tank fill level history using observations of the tank collected at different times. When we receive a new observation that is compatible with the last observation, we average new and old observations to derive a more accurate measurement of the tank fill percent. When we receive a new observation that differs from the last observation, we replace the fill percent of the tank with the new observation. When no observations are collected for a tank, we copy-forward the previously-seen measurement estimate. Therefore, for a given daily estimate, we use observations from the past to fill-in individual tanks that were not observed. There is a trade-off between how far in the past we extend this imputation and the tanks’ typical volatility. The error for the imputed measurement will become larger to reflect greater uncertainty in the estimate. Before any of the observations make it to this estimation stage, we employ several filters to our collection of observations to ensure that we only use the most-trusted observations. Filters we apply to observations include:

● Corrupted images. Sensor glitches, oversaturated readings, etc. can lead to unusable images.

● Images from certain satellites that suffer from very large georegistration errors, i.e., we are not certain enough that the image is of the location the image provider claims.

● Observations of tanks with relatively small dimensions. The nature of our observation is such that for small tanks, both our estimate of its dimensions as well as our estimate of its fill percent suffers relatively larger errors.

● Clouds. Using Orbital Insight’s proprietary Cloud Classifier, we are able to remove individual tank images that had clouds obscuring our view of the tank.

● Individual observations with low computer vision confidence scores. Our computer vision algorithms assign a score to each image which reflects how confident the algorithm is that it did well. This score can vary based on image noise, variations in image resolution, view geometry, etc.

● Observations for certain sun elevations. Because our method relies on shadows, for cases where the sun is too close to nadir, i.e., close to being directly above the tank, shadows become too small to reliably measure.

● Construction and destruction of the oil tanks. Orbital Insight has the world’s largest database of oil tanks but in order to inform our modeling approach and do backtesting, we use observations over the last six (6) years. Since that period of time, new tanks have been built and some have been demolished and/or replaced. We have created an extensive catalog containing each oil tank’s ‘lifetime.’ We use this lifetime catalog to

© Orbital Insight, Inc. v. 20181114

Page 5: Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation verview and Summary Orbital Insight Global Oil Storage leverages commercial satellite

ensure that we only use observations taken when the tank was actually present and to avoid accidentally interpreting a tank’s absence as a low fill percent.

After these filters have been applied, we create each daily estimate by taking the remaining observations (of tank fill percent) and computing a weighted average fill percent. We then multiply the average fill percent by the population’s storage capacity to arrive at the daily estimate of the oil volume in millions of barrels. As an example, in Figure 5, we show Orbital Insight’s estimate of oil volume in Cushing, OK, contrasted with an oil volume estimate from EIA. We also compute coherence and cross-correlation, not discussed here.

Figure 5: Comparison of Orbital Insight's and EIA's estimates of crude oil storage in Cushing, OK

Sampling Error Methodology Using our proprietary computer vision algorithms, we estimate the distribution of oil fill for each tank we observe, giving us a standard error for each tank measurement. This distribution is dependent on the satellite system and scene information. If a tank has not been observed for a few days, we propagate the error by increasing the standard error as a function of number of days since last measurement. Upon gaining a new measurement for an unobserved tank, we update the older measurement. Finally, at the region level, the standard error of individual tanks are aggregated together to obtain the region level standard error.

Scaled Estimate Methodology Scaled estimates are Orbital Insight’s volume estimates, renormalized to Government survey volumes (figure 6). This transformation aids in visualizing correlations in regions where Orbital Insight predictions have a lower scale than the Government survey (e.g. due to inclusion of non-FRT storage in Government survey volumes). We perform renormalization by scaling the Orbital Insight volume estimate for each day such that the Orbital Insight estimated and Government survey volumes have the same mean and variance in a trailing two-year window. This is a linear transformation and does not change the correlation between Orbital Insight and Government survey volumes. Because the transformation is calculated by matching mean and

© Orbital Insight, Inc. v. 20181114

Page 6: Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation verview and Summary Orbital Insight Global Oil Storage leverages commercial satellite

variance in a rolling window, the scaling transformation is slowly varying with time. This allows the transformation to automatically adjust for changes in the fraction of Government storage tracked by the Orbital Insight FRT storage volume (e.g. when new FRT storage is built).

Figure 6: Orbital Insight's Scaled Estimate and Error for PADD3 in the USA

Survey Correlation Metrics Methodology In regions where Government survey data are available, we provide a survey correlation metrics table that shows the statistical relationship between Orbital Insight estimated volumes and Government survey volumes (Figure 7). Correlation metrics are computed in a trailing N year window from "Latest Survey Date". "Correlation" is the Pearson correlation coefficient between the Government survey volume and Orbital Insight estimated volume (column volume.estimate) on each survey date. "Difference Correlation" is the Pearson correlation coefficient of the first difference of Orbital Insight and Government survey volumes. First differences are computed with Orbital Insight and Government survey volumes averaged over the last one, two and four survey dates. "Hit-Rate" is the fraction of weeks in which the first difference of Orbital Insight and Government survey volumes have the same sign.

© Orbital Insight, Inc. v. 20181114

Page 7: Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation verview and Summary Orbital Insight Global Oil Storage leverages commercial satellite

Figure 7: Orbital Insight's Survey Correlation Metrics for Cushing, OK (Survey: 2018-11-02)

Dataset Description Orbital Insight’s estimates of the world’s oil supply are limited to the measurement of floating-roof tanks (FRTs), excluding crude oil contained in fixed-roof tanks and other opaque sources such as underground storage. Therefore, we recommend that customers focus on the changes and trends in our time-series of available crude oil volume rather than drawing rigid conclusions on the absolute volume of oil. We assert that the crude oil volume fluctuations we observe in FRTs capture enough of the overall crude oil volume volatility to be able to produce a tradable, meaningful signal. The data we collect and use to infer from represents a sample, not a census of the population, and should be treated as such. The sample is limited by a number of factors, as outlined in this documentation. Noise should be expected and a degree of correlation/anti-correlation can be attributed to the random sample methodology. The datasets delivered to customers contains the fields described as follows:

date Date of estimate, measurements are typically between 10am-2pm (local time)

volume.estimate Estimate of the volume stored (millions of barrels) in the region of interest

smoothed.estimate 20-day simple moving average of volume.estimate.

© Orbital Insight, Inc. v. 20181114

Page 8: Orbital Insight Energy: Oil Storage v5.1 Methodologies ... · Methodologies & Data Documentation verview and Summary Orbital Insight Global Oil Storage leverages commercial satellite

volume.estimate.stderr Standard deviation of the volume.estimate in millions of barrels. We report the sampling errors associated with our measurements.

storage.capacity.estimate Estimate of total storage capacity for the tanks in the region of interest (millions of barrels)

total.available.tanks Total number of floating roof tanks in the region of interest

sampled.tanks Total number of individual tanks observed at least once in the last 15 days

truth_value_mb Ground truth, if available, from other sources such as Government surveys (eg, U.S. EIA)

An example view of the data (CSV) is shown below:

Dataset Delivery The output of the time series includes the estimated daily volume of oil (in millions of barrels) aggregated at the country, region or global level, and the corresponding date of the estimate. The time series can be viewed and downloaded through any of the following methods:

1. Orbital Insight’s Application Programming Interface (API): *.csv file 2. Orbital Insight’s Web Application: graphical User Interface (GUI), *.csv, *.json files 3. Using a Chicago Mercantile Exchange DataMine account: *.csv file

Individual tank level data (“raw data”) is also be available as an additional option.

© Orbital Insight, Inc. v. 20181114