Dual Contrastive Network for Few-Shot Remote Sensing Image Scene Classification

1. Problem Definition

The paper addresses Few-Shot Remote Sensing Image Scene Classification (FS-RSISC), a task aimed at classifying remote sensing images using only a few labeled samples per class. The authors identify two inherent challenges specific to remote sensing imagery that hinder existing methods:

• Small Inter-class Variances: Different scene categories (e.g., "Industrial" vs. "Residential") often look highly similar in terms of objects and arrangements, making them difficult to distinguish.
• Large Intra-class Variances: Images within the same category (e.g., "Palace") can vary significantly in background, lighting, and appearance, leading to high variability within a single class.

Most existing Few-Shot Learning (FSL) approaches rely on meta-learning, which the authors argue often underestimates the fundamental need for a robust, generalized feature extractor capable of handling these specific variance issues.

2. Methodology: Dual Contrastive Network (DCN)

The authors propose a transfer-based Dual Contrastive Network (DCN) that utilizes a pre-training stage followed by a meta-test stage. The core innovation is the integration of two auxiliary supervised contrastive learning branches during pre-training to learn complementary features:

A. Context-guided Contrastive Learning (CCL) Branch

• Goal: To address small inter-class variances by enhancing discriminative features that distinguish between different categories.
• Mechanism:
  • Condenser Network: A novel module designed to capture global context. It uses "Squeeze and Expand" operations (via 1x1 convolutions) to aggregate spatial information, re-weighting channels to emphasize scene-relevant information while suppressing irrelevant background noise.
  • Contrastive Learning: A supervised contrastive loss is applied to the condensed context features. This pulls features of the same class closer and pushes different classes apart in the embedding space, improving inter-class discriminability.

B. Detail-guided Contrastive Learning (DCL) Branch

• Goal: To address large intra-class variances by capturing invariant local details that remain consistent within a class despite background changes.
• Mechanism:
  • Smelter Network: Inspired by the smelting process, this module adaptively locates significant local regions. It performs Channel and Spatial Squeezing to generate attention maps, which are then used to highlight and enhance specific local detail features in the feature maps.
  • Feature Alignment: A self-attention based mechanism aligns feature maps between samples to accurately measure similarity.
  • Contrastive Learning: A contrastive loss is applied to these enhanced detail feature maps, forcing the model to focus on spatial details that are invariant within the same class.

C. Training and Inference

• Total Loss: The model is trained by minimizing a combination of Cross-Entropy loss (for classification) and the total Contrastive Loss ($L_{CL}$), which is the weighted sum of CCL and DCL losses.
• Inference: During the meta-test stage, the network freezes the feature encoder, Condenser, and Smelter networks. It generates prototypes using both the context features and the detail feature maps from the support set, combining their predictions for the final classification.

3. Key Contributions

1. Dual Contrastive Framework: The first work to explicitly combine Context-guided and Detail-guided contrastive learning branches to simultaneously tackle small inter-class and large intra-class variances in remote sensing.
2. Condenser Network: A specialized architecture that integrates and redistributes channel information to generate highly discriminative global context features.
3. Smelter Network: A mechanism that adaptively localizes and enhances significant local detail regions to capture invariant intra-class features.
4. Self-Attention Alignment: Introduction of a feature alignment mechanism to accurately measure similarity between detail feature maps.

4. Experimental Results

The DCN was evaluated on four benchmark remote sensing datasets: WHU-RS19, UC Merced, NWPU-RESISC45, and AID.

• Performance: The model achieved state-of-the-art (SOTA) performance across all datasets.
  • On NWPU-RESISC45 (5-way 5-shot), it outperformed the second-best method by 3.56%.
  • On AID (5-way 5-shot), it outperformed the second-best by 5.38%.
  • On small datasets (WHU-RS19 and UC Merced), it achieved the highest accuracy in both 1-shot and 5-shot settings.
• Ablation Studies:
  • Removing the Condenser Network or the Smelter Network significantly reduced performance, proving the necessity of both branches.
  • The dual-branch approach (DCN) consistently outperformed single-branch variants (CCL only or DCL only), demonstrating the value of feature complementarity.
• Visualization: Grad-CAM visualizations confirmed that the Condenser Network focuses on relevant scene objects (ignoring background noise), while the Smelter Network highlights specific local details, validating the theoretical design.

5. Significance

This paper makes a significant contribution to the field of remote sensing by shifting the focus from purely meta-learning strategies to robust feature representation learning. By explicitly modeling the dual nature of remote sensing variances (inter-class similarity and intra-class diversity), the DCN provides a more generalizable solution for few-shot classification. The proposed Condenser and Smelter networks offer new architectural paradigms for extracting global context and local details, which can be adapted to other computer vision tasks where class boundaries are blurry and intra-class diversity is high.