Data interpolation methods with the UNet-based model for weather forecast

Abstract Deep learning improves weather predictions, and most machine learning applications need data preparation, including interpolation. Since meteorological satellite collected data have several missing values, it is worth studying the interpolation in weather forecasting. This paper used a UNet-based model to evaluate 10 interpolation methods with different parameters on a short-term weather prediction task from the IEEE Big Data Competition 2021. Each strategy was evaluated using 3 groups of evaluation aspects, totaling 7 metrics. One of the specific issues explored in this research was reducing the influence of possible displacement in satellite images, which is often emphasized by exciting evaluation standards. After interpolation, some solutions showed that they could increase the model performance to some extent. Although there was no universal optimal method, interpolation using linear relationships performed rather well in most cases and produced the best results when all evaluation metrics were taken into account. However, the most effective method is time-consuming and requires a great number of calculations. In addition to interpolation with linear relation, computing the mean value of a limited region is beneficial and efficient. This study expects the conclusion to improve future weather prediction or meteorological data processing and to be expanded with other evaluation metrics to better assess a deep learning model’s effectiveness.