Active learning based on similarity level histogram and adaptive-scale sampling for very high resolution image...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Remote sensing image classification requires informative training samples.
• Existing methods may lack intra-class diversity due to spectral variations in remote sensing data.
• This can lead to suboptimal classification model performance.

Purpose of the Study:

• To propose an active learning method addressing intra-class diversity in remote sensing images.
• To improve classification accuracy for very high resolution remote sensing imagery.
• To mitigate sample imbalance issues in active learning.

Main Methods:

• Constructing similarity level histograms (SLH) for each class to capture intra-class diversity.
• Implementing an adaptive-scale sampling strategy to prevent sample imbalance.
• Utilizing active learning to select representative samples from SLH warehouses.

Main Results:

• The proposed method effectively considers intra-class diversity.
• Adaptive-scale sampling prevents over- or under-sampling issues.
• Experimental results demonstrate improved classification performance with limited training data.
• The algorithm is competitive with existing methods on public datasets.

Conclusions:

• The novel active learning approach enhances remote sensing image classification.
• Considering intra-class diversity and employing adaptive sampling are key to improving model performance.
• The method offers a competitive solution for efficient and accurate remote sensing image analysis.