Harmonizing Bi-Temporal and Multi-Sensor Satellite Imagery via a Vision–Language Framework for Comprehensive Disaster Assessment

Disaster assessment fundamentally requires reasoning across both temporal evolution and sensor heterogeneity. Yet existing methods typically address these dimensions in isolation, operating on single-temporal or single-modal satellite imagery and failing to capture the full scope of catastrophic changes. In this work, we present a vision–language framework that harmonizes bi-temporal and multi-sensor satellite observations within a unified architecture for comprehensive disaster recognition and damage mapping.
Our approach processes paired pre- and post-event optical and SAR images through modality-specific adapters that project heterogeneous inputs into a shared embedding space. A temporal comparison module, combining cross-time attention with residual feature differencing, captures disaster-induced transformations. To bridge the semantic gap, we align the fused visual representations with textual descriptions via contrastive learning, enabling zero-shot identification of disaster types and affected objects. Simultaneously, a language-conditioned decoder translates the same visual features into fine-grained segmentation masks for damaged regions such as collapsed buildings, flooded roads, and other referred targets.
We benchmark our framework against strong baselines including Siamese U-Net, SAM 2, and GroundingDINO–SAM 2, under both Optical–Optical and Optical–SAR scenarios. Extensive experiments demonstrate consistent superiority in zero-shot recognition (Accuracy, macro-F1) and semantic segmentation (mIoU, cIoU). Ablation studies isolate the contributions of temporal alignment, sensor adaptation, and language conditioning, confirming that each component is indispensable for handling heterogeneous satellite inputs. Notably, we find that linguistic guidance proves especially valuable for SAR imagery, where visual signatures alone are inherently ambiguous.
By jointly addressing temporal, sensor, and semantic heterogeneity within a single architecture, this work provides an effective, reproducible pathway toward next-generation disaster assessment systems. Code and annotations will be released to facilitate future research.