MetaCap: Adaptive Meta-Optimization for Few-Shot Remote Sensing Image Captioning

Remote sensing image captioning (RSIC) aims to translate visual content in remote sensing imagery into semantically coherent natural language descriptions. It has broad application value in fields such as disaster emergency response and cross-modal retrieval. Existing fully supervised approaches rely heavily on large-scale annotated data. However, due to factors such as scene diversity and long-tailed category distributions, many remote sensing scenarios are associated with limited annotated samples, resulting in few-shot RSIC. To address the few-shot RSIC challenge, we propose MetaCap, a novel cross-modal meta-captioning framework. Conventional metalearning methods suffer from optimization instability and cross-modal alignment difficulties when confronted with large inter-class variations in remote sensing imagery. To overcome these limitations, we propose a parameter-level adaptive optimization strategy that dynamically learns per-parameter update factors and weight decay factors during the inner loop, enabling precise adaptation to varying texture complexity across visual features. Furthermore, a visual-semantic consistency constraint and an inner-loop trajectory averaging mechanism are incorporated to effectively mitigate the semantic gap. Experimental results on the RSICD, UCM-Captions, and Sydney-Captions demonstrate the superior generalization ability of MetaCap.