Explainable spatio-temporal image segmentation for multivariate data : a case study on precipitation nowcasting

Abstract

Artificial Intelligence (AI) systems are often considered black boxes, making it difficult to understand the region of interest in the input for the classification and decision-making processes. This creates a challenge in convincing people who rely on traditional methods to trust the outputs of these systems. While there have been efforts to develop explainable AI systems for general machine learning problems, the same cannot be said for complex datasets such as the datasets used in precipitation nowcasting. In this paper, we adapt and apply existing Integrated gradients methods to understand and explain the predictions made by a deep learning model for precipitation nowcasting which takes a multi-image sequence as its input. We use the Meteonet dataset by MeteoFrance and form it into a spatiotemporal multivariate dataset consisting of rain radar, satellite, and wind images to predict rainfall 30 minutes ahead. We employ the U-Net model and evaluate the model’s explainability using Integrated gradients. The best performance achieved is an F1 Score of 0.76, consistent with other research using the same dataset. This lays a solid foundation for evaluating explainability. Our results show that this can not only explain the model predictions but also provide a visual representation of why the predictions were made, offering a potential solution for any spatio-temporal or multivariate dataset. Additionally, since this method provides a comparative analysis of how much each image in the input sequence contributed to the final result, we demonstrate how the same method can be further used for dimensionality reduction, enhancing the interpretability and efficiency of the predictive model.