RS-SSKD: Self-Supervision Equipped with Knowledge Distillation for Few-Shot Remote Sensing Scene Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Increasing availability of remote sensing (RS) imagery.
• Bottleneck in RS scene classification is the lack of ground truth samples, especially in unknown environments.
• Few-shot classification, a subfield of meta-learning, offers a solution by enabling knowledge extraction from limited data.

Purpose of the Study:

• To propose a novel method, RS-SSKD, for few-shot RS scene classification.
• To generate powerful representations for downstream meta-learners.
• To address challenges posed by insufficient training data in unknown environments.

Main Methods:

• A novel two-branch network architecture is proposed, utilizing three pairs of original-transformed images as input.
• Class Activation Maps (CAMs) are incorporated to guide the network in identifying category-specific regions.
• Self-knowledge distillation is employed to mitigate overfitting and enhance performance.

Main Results:

• The proposed RS-SSKD method demonstrates superior performance compared to current state-of-the-art approaches.
• Experiments were conducted on two challenging RS scene datasets: NWPU-RESISC45 and RSD46-WHU.
• Ablation studies confirmed the effectiveness of individual components and analyzed training time.

Conclusions:

• RS-SSKD effectively generates discriminative embeddings for few-shot RS scene classification.
• The method successfully overcomes limitations associated with insufficient training data.
• RS-SSKD represents a significant advancement in few-shot remote sensing scene classification.