Insights on the working principles of a CNN for forest height regression from single-pass InSAR data

Abstract

The increasing use of Artificial Intelligence (AI) models in Earth Observation (EO) applications, such as forest height estimation, has led to a growing need for explainable AI (XAI) methods. Despite their high accuracy, AI models are often criticized for their "black-box" nature, making it difficult to understand the inner decision-making process. In this study, we propose a multifaceted approach to XAI for a convolutional neural network (CNN)-based model that estimates forest height from TanDEM-X single-pass InSAR data. By combining domain knowledge, saliency maps, and feature importance analysis through exhaustive model permutations, we provide a comprehensive investigation of the network working principles. Our results suggests that the proposed model is implicitly capable of recognizing and compensating for the SAR acquisition geometry-related distortions. We find that the mean phase center height and its local variability represents the most informative predictor. We also find evidence that the interferometric coherence and the backscatter maps capture complementary but equally relevant views of the vegetation. This work contributes to advance the understanding of the model's inner workings, and targets the development of more transparent and trustworthy AI for EO applications, ultimately leading to improved accuracy and reliability in the estimation of forest parameters.