Reconstruction Guided Few-shot Network For Remote Sensing Image Classification

TL;DR

Few-shot remote sensing image classification is hindered by scarce labels and high land-cover variability. The authors introduce RGFS-Net, a reconstruction-guided framework that couples a masked image reconstruction task with a learnable bottleneck in an episodic meta-learning setting to learn robust, transferable features. The model optimizes a multi-task objective that combines four losses: $\mathcal{L}_{\text{proto}}$, $\mathcal{L}_{\text{triplet}}$, $\mathcal{L}_{\text{recon}}$, and $\mathcal{L}_{\text{var}}$, yielding $\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{proto}} + \alpha\,\mathcal{L}_{\text{var}} + \beta\,\mathcal{L}_{\text{triplet}} + \lambda\,\mathcal{L}_{\text{recon}}$. Experiments on EuroSat and PatternNet under 1-shot and 5-shot settings show RGFS-Net consistently surpasses state-of-the-art baselines, with ablations confirming the effectiveness of reconstruction, the bottleneck, and metric-learning components. The approach is simple, backbone-agnostic, and yields robust unseen-class generalization, making it practical for real-world remote sensing classification; code is available at the provided repository.

Abstract

Few-shot remote sensing image classification is challenging due to limited labeled samples and high variability in land-cover types. We propose a reconstruction-guided few-shot network (RGFS-Net) that enhances generalization to unseen classes while preserving consistency for seen categories. Our method incorporates a masked image reconstruction task, where parts of the input are occluded and reconstructed to encourage semantically rich feature learning. This auxiliary task strengthens spatial understanding and improves class discrimination under low-data settings. We evaluated the efficacy of EuroSAT and PatternNet datasets under 1-shot and 5-shot protocols, our approach consistently outperforms existing baselines. The proposed method is simple, effective, and compatible with standard backbones, offering a robust solution for few-shot remote sensing classification. Codes are available at https://github.com/stark0908/RGFS.

Figures (2)

• Figure 1: The model architecture of our proposed RGFS-Net. Here, we extract the image features from the pre-trained image encoder like (VGG-Net) which we have kept frozen during the training. We perform the reconstruction of masked patched image using learnable decoder under the supervision of reconstruction loss and mask loss. In order to better segregate the representaion space, we use the triplet loss along with the cross entropy loss.
• Figure 2: Comparison of model performance by varying the feature extractor backbone on the EuroSat dataset.