The Use of Observations in Learning Optimal Masks for California Temperature Prediction

Prediction of temperature is important for decision-making in a number of industries like agriculture, emergency management, healthcare, and more. However, long-term temperature forecasts on yearly to decadal timescales have limited skill. In an effort to improve this skill, many researchers have turned to machine learning methods for predicting climate variability and weather phenomena. We aim to quantify the accuracy of a yearly temperature prediction using machine learning at a lead time of 1-10 years. We combine machine learning with a technique called analog forecasting. Analog forecasting is the prediction of future states based on past states that had similar initial conditions, based on the premise that similar geophysical states evolve in similar ways. This method is intuitive for seasonal-to-decadal climate predictions and can outperform general circulation models (GCMs) in some cases. We use our interpretable neural network to learn a spatially-weighted mask that quantifies how important each spatial feature is for determining whether two states will evolve similarly. We then make a prediction for a specified lead time. We tested this method on temperature prediction in California, using the drought from 2012-2016 as a case study, since it is listed as one of the many causes of the catastrophic 2017 North Bay and Southern California wildfires. Our work demonstrates that a mask trained on ERA5 data is able to more successfully select useful analogs for accurate prediction from a library of models than one instead trained on the library of climate models alone.