XTR: An Explainable Transformer-Based Retracker for CryoSat-2 Altimetry

Published in IEEE Xplore: 08 June 2026
Authors: Alireza Dehghanpour, Danial Barazandeh, Gabriel Zachmann
ff GA XTR

Satellite radar altimetry is essential for monitoring the mass balance of polar ice sheets and their contribution to sea-level rise. However, over the dry-snow facies of Greenland and Antarctica, Ku-band pulses penetrate the firn, undergo volume scattering, and create time-dependent elevation biases. Traditional physical-model and threshold retrackers only partially compensate for this effect and usually require empirical post-processing based on waveform parameters such as backscatter and leading-edge width (LEW). Recently, deep-learning (DL) retrackers—especially convolutional neural networks (CNNs) trained on large ensembles of simulated waveforms—have achieved state-of-the-art accuracy by mapping echoes directly to surface range, yet they provide limited intrinsic interpretability and only indirect physical insight. This study presents XTR, an explainable Transformer-based retracker for CryoSat-2 low-resolution mode (LRM) waveforms. XTR couples a lightweight convolutional stem with multi-head self-attention to capture both local leading-edge structure and long-range waveform context, and uses mild waveform augmentation during training to improve robustness. We evaluate XTR using simulated waveforms and CryoSat-2 LRM crossover statistics over interior East Antarctica from 2020 to 2025 under a common retracker-only processing framework. On simulated data, XTR reduces the mean squared error (MSE) by approximately 35% relative to the classical threshold first-maximum retracker algorithm (TFMRA), although its error remains higher than that of the augmentation-matched AWI-ICENet1 CNN baseline. On real CryoSat-2 data, XTR lowers the crossover standard deviation by approximately 40% relative to AWI-ICENet1, while maintaining modest, region-dependent median offsets. Explainability analyses based on self-attention, SmoothGrad saliency, Integrated Gradients, and occlusion-based sensitivity tests indicate that XTR anchors its range estimate near the surface-dominated leading edge while using trailing-waveform structure as contextual information. Overall, our findings suggest that attention-based retrackers complement CNN-based schemes by offering competitive accuracy, intrinsic interpretability, and improved crossover stability in this retracker-only evaluation.