Baltimore Atlas: FreqWeaver Adapter for Semi-supervised Ultra-high Spatial Resolution Land Cover Classification

2025-06-18

AI Summary

Addressing the challenge of weakly supervised fine-grained land cover classification for 0.3-meter ultra-high-resolution remote sensing imagery—characterized by high annotation costs, large-scale variability, and poor transferability of large foundation models—this paper proposes a parameter-efficient semi-supervised segmentation framework. Methodologically, we build upon the SAM2 vision foundation model and introduce FreqWeaver, the first remote-sensing-specific frequency-domain adapter, which decouples high-frequency details via spectral feature modeling. Integrated with contrastive consistency regularization and pseudo-label self-training, our approach enables lightweight fine-tuning. With only a 5.96% parameter increase, it significantly improves structural consistency and boundary accuracy. On the Baltimore Atlas dataset, our method outperforms the best existing parameter-efficient approaches by +1.78% mIoU and surpasses state-of-the-art high-resolution segmentation methods by +3.44% mIoU, effectively alleviating the transfer bottleneck of large models in few-shot, multi-scale remote sensing scenarios.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Semi-Supervised Learning
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Vertical and domain-specific search

Abstract

Ultra-high Spatial Resolution Land Cover Classification is essential for fine-grained land cover analysis, yet it remains challenging due to the high cost of pixel-level annotations, significant scale variation, and the limited adaptability of large-scale vision models. Existing methods typically focus on 1-meter spatial resolution imagery and rely heavily on annotated data, whereas practical applications often require processing higher-resolution imagery under weak supervision. To address this, we propose a parameter-efficient semi-supervised segmentation framework for 0.3 m spatial resolution imagery, which leverages the knowledge of SAM2 and introduces a remote sensing-specific FreqWeaver Adapter to enhance fine-grained detail modeling while maintaining a lightweight design at only 5.96% of the total model parameters. By effectively leveraging unlabeled data and maintaining minimal parameter overhead, the proposed method delivers robust segmentation results with superior structural consistency, achieving a 1.78% improvement over existing parameter-efficient tuning strategies and a 3.44% gain compared to state-of-the-art high-resolution remote sensing segmentation approaches.

Problem

• High cost of pixel-level annotations for ultra-high resolution land cover classification
• Limited adaptability of large-scale vision models to high-resolution imagery
• Need for efficient semi-supervised methods to leverage unlabeled data

Innovation

• Parameter-efficient semi-supervised segmentation framework
• Remote sensing-specific FreqWeaver Adapter
• Lightweight design with minimal parameter overhead