What We Provide
We offer a trained deep learning algorithm for precise pixel-level contrail segmentation in satellite images.
Contrails, often overlooked, pose a significant challenge in our skies. Recent research reveals their surprising impact, they are responsible for 60% of aviation-related global warming and causing twice as much warming as direct CO2 emissions from aircraft. In the US alone, contrails blanket an area equivalent to Massachusetts and Connecticut combined.
Our mission is to help airlines and aircraft makers meet climate targets by reducing contrail
creation.
We contribute to a quick, low-cost, and high impact solution of contrail avoidance in flight
software planning tools by helping researchers improve their predictions.
We offer a trained deep learning algorithm for precise pixel-level contrail segmentation in satellite images.
Research indicates that diverting Only 1.7% of flights could reduce contrail climate damage by up to 59%. Contrail reduction also creates a new $9 billion carbon market forecasted to rise tenfold by 2030. Airlines have been enthusiastic about the subject as the benefits outweight costs.
Data scientists can use our model to analyze curated historical satellite images at specific locations and timestamps to identify the presence of contrails. This empowers data scientists working in the area of contrail prediction and avoidance by providing automatic contrail detection.
We serve the needs of data scientists at commercial aircraft manufacturers and flight software companies looking to predict contrail formation, evaluate and develop contrail avoidance systems.
We are using data from the paper: OpenContrails: Benchmarking Contrail Detection on GOES-16 ABI by Joe Yue-Hei Ng et. al. taken from the Kaggle Competition.
Satellite images from the GOES-16 ABI satellite which orbits the North and South American continent over April 2019-2020 (450GB):
Prediction labels containing binary labels identifying existence of contrails at pixel level from individual human labelers and the final ground truth labels:
Metadata shows the associated projection parameters such as the central meridian and timestamps of the satellite images.
The data has been processed by the researchers to downsample the negative instances by using:
False color image RGB channels were created by plotting brightness temperature contrasts using feature transformation of 3 bands.
Here is an example of a plot of one-single time step of an image (note there are 8 timesteps in the dataset):
Here is an example of contrails under one record moving and changing in shape over time:
TransUNet built on a hybrid Transfomer and UNet Architecture.
Combines the best of both worlds in learning global context and local information.
A model with Mean IoU Score of 0.7 to 0.8 is considered to be Very Good
We aimed to reduce the False positives to reduce the impact of misleading predictions
We Used UNet, Resnet101 and DeeplabV3 as baseline models, and a hybrid TransUnet using the original UNet architecture but with an additional Transformer Cycle
Model | Validation Mean IoU | Validation Dice Coefficient | False Positive Ratio | False Negative Ratio |
---|---|---|---|---|
MVP - TransUNet | 0.6997 | 0.7718 | 0.0988 | 0.1653 |
Baseline - UNet | 0.6951 | 0.7642 | 0.1550 | 0.1422 |
We hosted our model on Huggingface with test satellite images where we can detect contrails in
real-time
We are a team of graduate students in UC Berkeley's Master of Information and Data Science (MIDS)
Program.
We have expertise in computer vision, data engineering/processing, and user/domain
research.
We would like to express our sincere thanks to the following people who were integral to our project: