Notes on Carbon Plan

September 7, 2023 · 2 min read

Notes on an article from Carbon Plan.

Cover

Earth Climate Models(ECMs) suffer from inaccuracy because of the resolution, a small pixel might represent 100-200km depending on the location. This pixel is given the same attributes like altitude, temperature etc., eventhough in reality the there might be huge altitude differences within that pixel, which influences the weather in a different way. Simply said, Estimation is not fine enough Why?
Because Earth Climate Models are designed to capture global patterns not local climate patterns. why? Maybe data samples are available at that resolution. This creates the need to downscale the data.

Downscaling pipeline takes the global datasets and historical data to create downscaled datasets aka data at the local level. One can then use different models to predict precipitation, floods, temperature etc at the local level.
The type of downscaling algorithm produces different downscaling datasets, which makes the local weather prediction a bit unreliable.
- Questions?
- what about using all downscaling datasets and averaging them?
- Does the historical datasets contain the samples related to the current crisis?

downscaling

algs

Training the downscaling algorithms:
- Simply put, we need a black box(if its a neural network) that converts climate data at macro level to micro level.
- Carbon plan does this by training a model that uses historical macro climate data and converts this to climate data at micro level. The Groundtruth data at micro level are historical observations.
- Once the model is accurate enough in conversion, they use climate data projections at macro level as an input to the trained model and the model outputs the climate projections at the micro level.