Uncertainty-aware AI boosts remote sensing image-text retrieval robustness
New evidential learning method handles degraded images and vocabulary mismatches reliably.
A team of researchers (Wang, Wang, Li, Li, Liu, Ban) has introduced ELC, a novel uncertainty-aware method for cross-modal remote sensing image-text retrieval (CMRSITR). Existing approaches treat every query with full confidence, failing under real-world conditions like image degradations from sensors or atmosphere, and text-side vocabulary heterogeneity. ELC tackles this by employing evidential learning (EDL) during training to model inter-modal correspondences as Dirichlet distributions, enabling the system to quantify uncertainty for each query.
Key innovations include Uncertainty-Correctness Alignment Learning (UCL), which aligns estimated uncertainty with actual retrieval correctness—low uncertainty for correct retrievals, high for incorrect ones. Additionally, Intra-Modal Relationship Learning (RL) distills similarity structure from pretrained mentors to make the Dirichlet distributions more discriminative. At test time, low-uncertainty queries are returned directly, while high-uncertainty ones are refined via RS-aware test-time augmentation (RS-TTA). Experiments confirm ELC achieves competitive accuracy and significantly better robustness under sensor- and atmosphere-related perturbations and vocabulary mismatches.
- Uses evidential learning to capture uncertainty in cross-modal image-text retrieval
- Introduces uncertainty-correctness alignment and intra-modal relationship learning
- Outperforms state-of-the-art under sensor/atmosphere degradations and vocabulary heterogeneity
Why It Matters
Enables more reliable AI for satellite imagery analysis, search and rescue, and environmental monitoring.