Text-embedded domain-generalized state space model for the visual question answering of SAR images

Official URL: https://dx.doi.org/10.1109/LGRS.2026.3726964

Visual question answering (VQA) models are gaining popularity in remote sensing, especially in the context of synthetic aperture radar (SAR) images; yet they are being developed under the assumption that training and testing data are drawn from the same distribution, which is unrealistic in real-world scenarios where domain shifts frequently occur. In this letter, this task is investigated from a domain generalization (DG) perspective, where we hypothesize that integrating publicly available country-specific geographical knowledge into a model will boost its question answering performance, assuming it has access to the country of origin during inference. To this end, a text-embedded state space architecture is proposed to fuse geographical text with SAR images. Extensive experiments and ablation studies show that it achieves state-of-the-art performance. The code is publicly available at github.com/Elman295/TESSA.