A Training-Free Paradigm for Data-Scarce Maritime Scene Classification Using Vision-Language Models

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Computer Vision

Background:

• Maritime Domain Awareness (MDA) heavily relies on high-resolution optical spaceborne sensor data.
• Traditional supervised deep learning for MDA faces significant challenges due to the need for exhaustively annotated datasets.
• Data scarcity severely degrades the performance of conventional models, hindering real-time applications.

Purpose of the Study:

• To develop a training-free inference paradigm for MDA that overcomes data scarcity limitations.
• To leverage the pre-trained knowledge of Large Vision-Language Models (VLMs) for enhanced MDA.
• To bridge the perspective gap between natural images and top-down optical sensor imagery.

Main Methods:

• Introduction of a Domain Knowledge-Enhanced In-Context Learning (DK-ICL) framework.
• Coupling DK-ICL with a Macro-Topological Chain-of-Thought (MT-CoT) strategy.
• Translating expert remote sensing heuristics into a step-by-step reasoning pipeline for VLMs.

Main Results:

• MT-CoT augmented VLMs achieved over 38% higher F1-score compared to traditional models under data scarcity.
• The zero-gradient approach demonstrated robust generalization on unannotated, out-of-distribution coastal clutters.
• Performance parity was achieved with data-heavy networks using 50 times less data volume.

Conclusions:

• The proposed training-free paradigm offers a resource-efficient foundation for next-generation intelligent maritime sensing networks.
• This approach substitutes massive human annotation and GPU optimization with scalable logical deduction.
• The framework significantly enhances MDA capabilities in data-limited environments.