HSLabeling: Toward Efficient Labeling for Large-Scale Remote Sensing Image Segmentation With Hybrid Sparse Labeling.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Dense pixel-wise labeling of large-scale remote sensing images (RSI) is costly and time-consuming.
• Existing sparse label methods often use a single label type, limiting segmentation performance on complex land covers.
• Different land cover types benefit from different sparse label representations (e.g., points, scribbles, blocks).

Purpose of the Study:

• To develop a novel framework, Hybrid Sparse Labeling (HSLabeling), for efficient and accurate large-scale RSI segmentation using adaptive sparse labels.
• To address the limitations of single-type sparse labeling in capturing diverse land cover characteristics.
• To reduce labeling costs while maintaining high segmentation performance.

Main Methods:

• Proposed a label segmentation contribution information estimation module to assess the value of different sparse labels based on land cover diversity and shape.
• Developed an Optimal Hybrid Labeling Strategy (OHLS) to adaptively assign optimal sparse label types to different land covers.
• Formulated label assignment as an optimization problem balancing contribution and cost, solved efficiently using a greedy algorithm.

Main Results:

• HSLabeling achieved near fully supervised performance on three large-scale RSI datasets with minimal labeling effort.
• The framework adaptively selected optimal hybrid sparse labels for varied land covers.
• Compared to single sparse label methods, HSLabeling achieved similar performance with significantly lower labeling costs.

Conclusions:

• HSLabeling offers an effective solution for cost-efficient and high-performance segmentation of large-scale remote sensing images.
• The adaptive selection of hybrid sparse labels is crucial for handling complex land cover segmentation.
• This approach significantly reduces the practical barriers associated with acquiring labeled data for RSI analysis.