Joint inter-seasonal forecasts with deep multitask learning

Abstract:

Seasonal and sub-seasonal forecasts have seen significant progress in recent years with the aid of machine learning. To work effectively, machine learning models heavily rely on large amounts of high-quality labeled data, which are expensive or even impossible to build for certain applications. Deep convolutional neural networks are machine learning models that are capable of modeling complex spatial patterns in climate data but are even more data hungry than traditional machine learning methods, mainly due to the larger number of parameters that need to be learned. In this work, we investigate deep multitask learning (MTL) to jointly provide inter-seasonal forecasts for multiple climate variables. MTL is a machine learning paradigm in which multiple regressors are trained jointly while sharing commonalities among them to obtain more accurate regressors. MTL has been shown to reduce the “sample complexity” of the learning process, that is, the amount of labeled data needed to train a machine learning model. More specifically, we train multiple deep convolutional neural networks, one per seasonal forecast of a specific climate variable, in a shared framework. Therefore, allowing for information to flow from one task to the other. We show MTL results on seasonal climate predictions in the Western United States for temperature, precipitation, and snowpack.