A Unified Framework for Efficient Remote Sensing Visual Question Answering: Adapting Dual, Hybrid, and Encoder-Decoder Architectures

TL;DR

A PEFT framework injecting lightweight adapters into three VLM architectures for remote sensing VQA, showing hybrid FLAVA achieves the best balance.

Problem

Direct application of general-domain VQA models to remote sensing is hindered by domain shift and high computational cost of full fine-tuning.

Approach

Inject lightweight bottleneck adapters into attention and MLP layers of frozen CLIP, BLIP, and FLAVA backbones, training less than 5% of parameters on RSVQA-x dataset.

Results & Contribution

All adapted models converge; hybrid FLAVA offers superior multimodal reasoning and retrieval. Establishes a new baseline for resource-efficient VQA in disaster assessment and urban monitoring.