Spatial coherence-based batch-mode active learning for remote sensing image classification.

Area of Science:

• Remote Sensing
• Machine Learning
• Computer Vision

Background:

• Batch-mode active learning (AL) selects training samples using uncertainty and diversity criteria.
• Conventional AL methods neglect spatial coherence in remote sensing, increasing costs.
• Spatial coherence is crucial for efficient sample selection in remote sensing data.

Purpose of the Study:

• To propose a novel spatial coherence-based batch-mode active learning (AL) method for remote sensing applications.
• To address the limitations of existing AL methods that ignore spatial relationships between training samples.
• To reduce the cost of training sample selection in remote sensing by incorporating spatial coherence.

Main Methods:

• Utilizes mean shift clustering for the diversity criterion, allowing variable query numbers per iteration.
• Represents spatial coherence using a two-level segmentation map for automatic labeling of new queries.
• Introduces a new merging strategy for mean shift segmentation to ensure stable and accurate segmentation maps.

Main Results:

• The proposed method demonstrates effectiveness in reducing costs associated with training sample selection.
• Experimental results on two real remote sensing datasets validate the proposed techniques.
• The method outperforms other state-of-the-art methods in terms of efficiency and accuracy.

Conclusions:

• The developed spatial coherence-based batch-mode AL method is effective for remote sensing.
• Incorporating spatial coherence significantly improves the efficiency and reduces the cost of AL in remote sensing.
• The novel merging strategy for mean shift segmentation enhances the quality of spatial coherence representation.