ClimateHack.AI 2022
ClimateHack.AI 2022 was a satellite imagery generation challenge that ran between January & March 2022.
Loading...
ClimateHack.AI 2022 was a satellite imagery generation challenge that ran between January & March 2022.
Join the official ClimateHack.AI Discord server to become part of an international community of AI enthusiasts and receive competition announcements. 🚀
Climate Hack.AI 2022 was a collaborative initiative between the student communities of 25 universities leading in CS and AI from across the United States, the United Kingdom and Canada to take a lead in the fight against climate change using machine learning.
Participants had two months to apply cutting-edge machine learning techniques in order to develop the best satellite imagery prediction algorithm for use in solar photovoltaic output forecasting applications.
ClimateHack.AI 2022 was open to students, no matter their level of machine learning experience, who were enrolled at the following universities: Bristol, Caltech, Cambridge, Carnegie Mellon, Columbia, Cornell, Edinburgh, Georgia Tech, Glasgow, Harvard, Illinois Urbana-Champaign, Imperial College London, Manchester, Michigan, MIT, Oxford, Princeton, St Andrews, Stanford, Toronto, University College London (UCL), UC Berkeley, UCLA, Warwick and Waterloo.
The winning teams of ClimateHack.AI 2022 came from the following institutions:
The first-place team's model was the best model Open Climate Fix had seen for cloud movement tracking, and Open Climate Fix's live production solar power forecasting service is being used by the National Grid Electricity System Operator in Great Britain to cut emissions potentially by up to 100 kilotonnes per year.
We could not have run ClimateHack.AI 2022 without the generous support of Newcross Healthcare.
Archived competition information
The ultimate end-goal is to collaborate to build the world's best near-term forecasting system for solar electricity generation, which would allow electricity system operators around the world to better schedule their grids, saving potentially ~50 million tonnes of CO2e a year by 2030 if deployed worldwide.
The key to this lies in developing better satellite imagery nowcasting techniques, which would allow better cloud coverage predictions, amongst other things. This is where you come in.
In order to account for the variability of solar photovoltaic (PV) power production, the National Grid Electricity System Operator (ESO) schedules a spinning reserve of natural gas generators, which can take hours to ramp up from a cold start, to operate below their maximum capacity so that there is headroom on the grid that can ramp up rapidly to make up for any shortfalls.
Not only is this expensive, but it contributes to ~100 kilotonnes in excess carbon emissions each year in Great Britain alone. As such, better solar PV forecasting techniques would allow the National Grid ESO to cut their spinning reserve, thereby reducing emissions and helping to improve the deployability of cheaper, greener solar power.
Cloud coverage (especially in areas with variable meteorological environments such as the United Kingdom) can have a outsized impact on solar photovoltaic power yields. By incorporating satellte imagery into near-term forecasting (or "nowcasting") models for solar power generation, it is possible to significantly improve the minute-to-minute accuracy of machine learning-based solar PV models beyond relying on numerical weather predictions alone.
Take a look at this video animation from Open Climate Fix to see this effect in practice:
For more background and motivation, check out this Wired article about the solar electricity forecasting work of our partner, Open Climate Fix.
Your challenge—should you choose to accept it—is to predict the next two hours of satellite imagery from the previous hour of satellite imagery over quadruple the area better than the current state of the art before submissions close on Wed 16th March (23:59 GMT).
Task: From twelve 128×128-pixel images taken five minutes apart (one hour of data), predict the next two hours of satellite imagery for the smaller central 64×64-pixel region. The aim is for solutions to produce accurate, sharp, non-blurry images close to the ground truth.
Input: an hour of satellite imagery for a 128×128-pixel region (12 timesteps × 128 pixels × 128 pixels), as well as the geospatial positions of the images (which may be useful to feed into your model).
Output: the next two hours of satellite imagery for the 64×64-pixel area at the centre of the input region (24 timesteps × 64 pixels × 64 pixels).
Note that this means the spatial extent of the input is four times larger than that of the output!
For the loss function and scoring metric, Open Climate Fix recommend using the multi-scale structural similarity index measure (MS-SSIM). In their experience, MSE tends to encourage models to produce overly blurry predictions, so MS-SSIM is a better option. They have a PyTorch implementation of a differentiable MS-SSIM in their GitHub repository.
Having said that, you are completely free to experiment with different loss metrics (e.g. MAE, PSRN, etc) or even a combination thereof; while MS-SSIM is currently used on the leaderboard, all interesting solutions that output clear, accurate image sequences are welcome! Producing sharp images useful in cloud coverage and solar photovoltaic output forecasting should be prioritised over purely maximising submissions' MS-SSIM scores.
The satellite imagery dataset includes data from all hours of the day. It is recommended that you only select "daylight" hours, as the submissions will be tested on "daylight" hours only. "Daylight" hours are defined as as hours where the sun is at least 10 degrees above the horizon, as measured from the centre of the 128×128 pixel input image. The angle of the sun can be computed using pvlib.solarposition.get_solarposition.
Given the satellite images are 1,843×891 pixels, you can get a huge number of 128×128 training examples by randomly selecting 128×128 crops from the satellite imagery. You just want to make sure that the entire temporal extent of each example is in "daylight", i.e. the sun is at least 10 degrees above the horizon as measured from the centre of each 128×128 crop.
Open Climate Fix have provided "high resolution visible" satellite imagery taken over the UK and north-western Europe from a geostationary orbit between January 2020 and November 2021. It was sourced from EUMETSAT's Spinning Enhanced Visible and InfraRed Imager Rapid Scanning Service. The individual images are 1,843×891 and have a spatial resolution of about 2-3 km (decreasing south to north). The satellite produces images every five minutes, so an hour of satellite imagery consists of 12 timesteps.
Check the dataset out on Google Cloud Public Datasets.
To get started, download and extract the example project from GitHub. In addition to containing more information about the dataset and the challenge, the Jupyter notebook in training.ipynb guides you through training a basic model and making your first Climate Hack.AI submission.
Also, make sure to check out Open Climate Fix's notebook, which provides a more in-depth introduction to reading the satellite data, as well as more technical information about the dataset.
For this challenge, we evaluate submissions inside a lightweight Linux-based virtual machine with 4 vCPUs and 6GB RAM. Submissions may be no larger than 4GB in size.
The Python 3.9 packages installed in the evaluation environment include the following:
cachetools (5.0.0),
dm-reverb (0.6.1),
dm-tree (0.1.6),
flatbuffers (2.0),
gym (0.21.0),
joblib (1.1.0),
keras (2.8.0),
Keras-Preprocessing (1.1.2),
numba (0.55.1),
numpy (1.21.5),
opencv-contrib-python-headless (4.5.5.62),
opt-einsum (3.3.0),
einops (0.4.0),
perceiver-pytorch (0.8.3),
pandas (1.4.0),
Pillow (9.0.1),
scikit-learn (1.0.2),
scipy (1.7.3),
tensorboard (2.8.0),
tensorflow (2.8.0),
tf-agents (0.7.1),
torch (1.10.2+cpu),
torchaudio (0.10.2+cpu),
torchvision (0.11.3+cpu),
wrapt (1.13.3)
If there is a package you want to use that is not installed in the evaluation environment, reach out to us on Discord and we will try our best to accommodate any reasonable requests!
In-person finals will be held at Harvard University and University College London from Thursday 24th March until Saturday 26th March 2022. Look out for an email from us when the qualifying round closes if you expect to be in the top three for your academic institution!