Exploring the Impact of SAR Images on Wave Height Prediction Models: A Study Using XAI

This paper presents an application of explainable AI (XAI) techniques to a Convolutional Long Short-Term Memory (ConvLSTM) wave height prediction model based on Synthetic Aperture Radar (SAR) imagery and satellite altimetry data. Using XAI methods such as occlusion sensitivity and integrated gradients, we explore the spatial and temporal features most critical for accurate wave height predictions. The XAI analysis reveals that the model heavily relies on early SAR images to capture essential sea surface roughness patterns, while later images primarily serve to refine predictions. These findings enhance the transparency and interpretability of wave height models, providing valuable insights for real-world applications such as coastal management and disaster preparedness, where understanding model decision-making is essential for trustworthy predictions.