Optimizing Deep Learning for Satellite Hyperspectral Data: an xAI-Driven Approach to Hyperparameter Selection

The growing availability of spaceborne Hyperspectral Imaging (HSI) missions combined with advancements in Deep Learning (DL), offers significant potential for global environmental mapping. However, most DL methods are tailored to airborne HSI, making it challenging to adapt them for optimal performance with satellite data. To solve this issue, this study explores the use of explainable Artificial Intelligence (xAI) to adapt existing DL architectures to spaceborne HSI. In particular, the best hyperparameter selection is carried out by evaluating the consistency of the explanations across different model instances using xAI Integrated Gradients method. Experiments were conducted using two cutting-edge DL models commonly used for airborne HSI: an attention-based Vision Transformer (ViT) and a standard 2D-Convolutional Neural Network (CNN). Results from a crop-type mapping task using Hyperspectral Precursor and Application Mission (PRISMA) satellite data demonstrate the effectiveness of the proposed approach in facilitating the optimal hyperparameter selection, i.e., the one able to maximize the macro Fscore obtained on the test set.