Cross-Region Few-Shot Remote Sensing Image Captioning via Adaptive Vision-Language Feature Fusion

Abstract

Remote Sensing Image Captioning (RSIC) enables automated interpretation of aerial imagery by converting complex visual scenes into coherent natural language descriptions. A key challenge in RSIC is the scarcity of annotated data and the significant domain shifts across geographic regions. Models trained on specific regional features often degrade in performance when applied to visually distinct landscapes such as agricultural or coastal areas. To address this, we propose the Adaptive Vision-Language Feature Fusion (AVLF) network, a few-shot learning framework designed to achieve robust cross-region transfer with minimal data. The AVLF framework bridges the semantic gap between visual and linguistic representations through an adaptive gating mechanism that dynamically balances visual and language features during caption generation. Extensive experiments on cross-region splits of multiple remote sensing datasets demonstrate that AVLF achieves state-of-the-art performance, maintains high captioning quality with limited support sets, generalizes effectively to unseen semantic categories, and incurs minimal computational overhead. Feature space visualizations show well-separated class distributions, while attention maps confirm that the model focuses on semantically relevant geographic objects. Ablation studies further highlight the importance of the adaptive fusion strategy in overcoming domain discrepancies and enhancing few-shot learning capability.