Oil storage tanks play an important role in the global economy. Crude oil is stored in tanks at many points between extraction and sale. Storage tanks are also used by nations to stockpile oil reserves. The volume of oil in storage is an important economic indicator. It indicates which oil producing nations are increasing or decreasing production and gives a window into the global demand for energy. At the same time, oil storage information is not transparent. Nations may hide information about oil production, consumption and storage for economic or military reasons. For this reason, companies like Planet and Orbital Insight have made a business of collecting satellite imagery of oil storage tanks and estimating reserve volumes.
Volume estimation is made possible by the design of storage tanks. The tanks have a floating head that sits on top of the oil. This prevents flammable vapors from building up. It also creates a distinctive pair of two crescent shadows:
These shadows can be extracted from the image using computer vision algorithms.
The relative areas of the shadows can be used to estimate the volume of the tank.
This project consists of four parts:
-
- Dataset Creation
-
- Tank Object Detection
-
- Shadow Extraction and Volume Estimation
-
- Full Processing Pipeline
Dataset creation is in progress. 100 4800x4800 RGB images of tank containing industrial regions were collected from Google Earth. Each large image was split into 100 512x512 tiles, for a total of 10,000 512x512 tiles. Image data with bounding box labels for tiles are available on Kaggle
A RetinaNet object detection model was trained on the dataset of 9000 images to detect and localize tanks in images, detailed in the Object Detection notebook. The model shows the ability to accurately locate tank images, but struggles to generate precise bounding boxes. This is likely due to the limited size of the dataset. 9000 images is small for object detection (compared to say 200,000+ for COCO), and many of the images don't contain tanks. A future direction would be to use a model pre-trained on COCO or another object detection dataset to see if that boosts performance for the tank identification task.
Identified tanks will be run through a computer vision algorithm to extract the crescent shadows and estimate tank volumes. The full algorithm is detailed in the Shadow Extraction Algorithm notebook. In short, the algorithm creates an enhanced version of the tank image using challens from RGB and LAB color space. Shadows are extracted from the enhanced image using thresholding. The thresholded image is cleaned up using morphological operations. Tank shadows are extracted from the clean image and used for volume estimation.
The algorithms shown here present the basis of a data pipeline for estimating oil tank volumes. A full pipeline would run something like this:
- Break large satellite images into 512x512 patches
- Run patches through the object detection network
- Use bounding boxes from detected tanks to run the shadow extraction and volume estimation algorithm
- Reassemble image patches into the full size image with volume estimations added