A Progressive Semi-Distillation Model for Dual-Source Remote Sensing Image Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Dual-source remote sensing image classification is a growing research area.
• Limited labeled samples pose a significant challenge for accurately classifying dual-source images.
• Existing methods struggle to effectively leverage dual-source information with insufficient data.

Purpose of the Study:

• To propose a progressive semi-distillation model (PSDM) for dual-source remote sensing image classification.
• To address the challenge of insufficient labeled samples in dual-source image classification.
• To enhance the accuracy, efficiency, and robustness of classification models under data scarcity.

Main Methods:

• A progressive semi-distillation model (PSDM) framework incorporating a rookie teacher network (RTN), teaching assistant system (TAS), and student grouping network (SGN).
• The RTN-SGN structure is employed to expand samples and compress feature space, mitigating the insufficient sample problem.
• The TAS gradually guides the SGN from easy to difficult samples, improving training and performance.

Main Results:

• The proposed PSDM effectively handles insufficient labeled samples in dual-source remote sensing image classification.
• The SGN, with its cooperation and correction mechanisms, outperforms the RTN, demonstrating the effectiveness of semi-distillation.
• Experimental results validate the accuracy, efficiency, and robustness of the PSDM.

Conclusions:

• The PSDM offers an effective solution for dual-source remote sensing image classification with limited data.
• The developed framework successfully overcomes the limitations of insufficient samples.
• The model achieves superior performance and robustness, making it a valuable contribution to the field.