SUMMIT: A SAR foundation model with multiple auxiliary tasks enhanced intrinsic characteristics

Synthetic Aperture Radar (SAR) is a crucial tool in remote sensing, yet existing deep learning methods are primarily limited in visual representation, neglecting the intrinsic characteristics of SAR and the need for strong generalization across multiple tasks. To address this, we propose SUMMIT (SAR foUndational Model with Multiple auxiliary tasks enhanced Intrinsic characterisTics), a foundational model tailored for SAR image understanding. SUMMIT is pre-trained on the Multi-sensor SAR Image Dataset (MuSID), which contains over 560,000 SAR images. To enhance its feature extraction capability, we introduce a masked image modeling (MIM) framework with self-supervised auxiliary tasks (SSATs): (1) MIM for learning robust structural representations, (2) self-supervised denoising to improve the model’s noise resistance, and (3) space scattering feature enhancement to preserve geometric consistency. Furthermore, we design an auxiliary task coordination module (ATCM) to balance these tasks and ensure effective feature fusion. The resulting self-supervised framework enables SUMMIT to integrate deep learning with SAR’s physical attributes effectively. Extensive experiments across seven datasets and three downstream tasks demonstrate that SUMMIT achieves state-of-the-art performance, particularly in SAR classification, detection, and segmentation. Code and pre-trained model of the proposed SUMMIT will be available at https://github.com/Yunsans/SUMMIT . • This paper explores foundation models with SAR intrinsic fusion. • We build a large dataset with over 560k images for pre-training. • A multi-auxiliary task joint self-supervised framework is proposed. • Denosing and scattering feature enhancement are designed as auxiliary tasks. • All auxiliary tasks are integrated into an end-to-end framework.