Few-shot remote sensing scene classification based on multi subband deep feature fusion.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in object recognition but struggle with limited labeled data in remote sensing (RS).
• Insufficient labeled samples hinder the practical application of deep learning in RS image processing.
• Existing methods often fail to effectively utilize the rich information within RS images under data scarcity.

Purpose of the Study:

• To develop an effective method for small sample remote sensing image classification.
• To address the challenge of insufficient labeled data in RS image analysis.
• To improve the discrimination of easily confused categories in RS images.

Main Methods:

• Utilizing pre-trained deep CNNs and discrete wavelet transform (DWT) for deep feature extraction from RS images.
• Proposing a modified discriminant correlation analysis (DCA) to enhance feature discrimination based on between-class distance coefficients.
• Integrating deep features from various frequency bands using the proposed DCA approach.

Main Results:

• The proposed method effectively fuses multi-level deep features from RS images.
• Achieved low-dimensional features with strong discriminative power.
• Demonstrated outstanding performance on four benchmark datasets, especially with one or two training samples per class.

Conclusions:

• The discrete wavelet-based multi-level deep feature fusion method significantly enhances RS image classification accuracy with limited samples.
• The modified DCA effectively distinguishes between similar categories, overcoming a key limitation in RS data analysis.
• This approach offers a promising solution for practical RS image classification tasks where labeled data is scarce.