CPRS

Coastal-Port Remote Sensing · The data behind FBA

CPRS is the data behind FBA. It brings together remote sensing image–caption pairs and instruction dialogues for coastal harbor understanding.

This page is a preview with examples to explore. The full data and preparation notes will follow after paper acceptance.

先放一些样例，方便大家看看。论文录用后，会陆续补上完整数据和整理方法。

What’s inside

About 810K supervision records, organized in three layers:

Layer	Records	Focus
RS-Anchor	569,853 image-caption pairs	Overhead imagery and remote sensing language
Bridging-Conv	187,296 SFT records	Related scenes across RGB, SAR, PAN, and NIR
Scenario-EG	53,000 SFT records	Harbor questions, spatial reasoning, localization, and reports

Browse some examples

We’ve picked eight scenes, two per sensor, for the interactive gallery. Pick an image, switch between questions, and explore the visual clues behind each answer.

RGB	SAR
PAN	NIR

You can find the images in the GitHub examples, or visit HarborEval for evaluation examples.

Cite the paper

@article{zong2026fba,
  title   = {Filling Before Advancing: Capability-Gap-Driven Post-Training for Scenario-Specialized Remote Sensing MLLMs},
  author  = {Zong, Yuheng and Wang, Minghua and Zhao, Xin and Zhan, Zhi-Hui and Plaza, Antonio and Benediktsson, Jon Atli},
  journal = {arXiv preprint arXiv:2607.22205},
  year    = {2026},
  doi     = {10.48550/arXiv.2607.22205},
  url     = {https://arxiv.org/abs/2607.22205}
}