Optimization of Deep Learning Precipitation Models Using Categorical Binary Metrics

This work introduces a methodology for optimizing neural network models using a combination of continuous and categorical binary indices in the context of precipitation forecasting. Probability of detection or false alarm rate are popular metrics used in the verification of precipitation models. However, machine learning models trained using gradient descent cannot be optimized based on these metrics, as they are not differentiable. We propose an alternative formulation for these categorical indices that are differentiable and we demonstrate how they can be used to optimize the skill of precipitation neural network models defined as a multi-objective optimization problem. To our knowledge, this is the first proposal of a methodology for optimizing weather neural network models based on categorical indices. Plain Language Summary Deep neural networks have recently demonstrated great versatility and an unprecedented capacity to model complex problems. In weather modeling, these algorithms have been applied to solve different problems. This is a promising area of research, given the availability of large volumes of weather data and increasingly powerful computers. Neural network models can learn to solve problems based on a metric, which the model tries to optimize. However, the quality of weather models is measured using a large variety of metrics, which can be a challenge when choosing which metric the model should optimize. In the case of precipitation, categorical binary metrics are a popular choice to asses the quality of a model. These metrics reduce precipitation to a ’yes’ or ’no’ event and the results of the predicting model can be compared with the actual observations. This method is simple, yet powerful and a large number of indices and statistics have been developed to assess different aspects of the quality of precipitation models. As precipitation models are commonly assessed using these categorical binary metrics, it would be very convenient to optimize models based on them. Unfortunately, the mathematical nature of these metrics makes them unsuitable for optimizing deep learning models. In this work we present an alternative formulation for these categorical binary indices which can be used to train models. We demonstrate how a deep learning model can be trained to generate better quality precipitation data.