Scholarlink

Few-shot learning (FSL) provides a practical solution for remote sensing applications with limited labeled data, yet its effectiveness on medium-resolution satellite imagery remains underexplored. This study investigates FSL methods for classifying Maasai bomas using PlanetScope imagery at 3.0– \(3.5~\text {m}\) ground sampling distance. To replicate operational constraints, we adopt an episodic evaluation framework where a small number of support samples per class and corresponding query samples are randomly selected from the dataset. We benchmark a state-of-the-art few-shot learning approach that integrate transfer learning and meta-learning strategies. For feature extraction, we evaluated Vision Transformers pretrained through DINO v1, DINO v2, and DINO v3, as well as convolutional architectures including ConvNeXt and Swin v2. Each backbone was evaluated with a Prototypical Network framework to measure its ability to generate discriminative embeddings under limited supervision. Furthermore, we compared models pretrained on out-of-domain datasets with those trained using in-domain satellite imagery to examine how pretraining domain influences generalization and adaptation in remote sensing tasks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Benchmarking Few-Shot Methods for Rare Target Classification in PlanetScope Imagery

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Keli Cheng,
  • Weipeng Wu

摘要

Few-shot learning (FSL) provides a practical solution for remote sensing applications with limited labeled data, yet its effectiveness on medium-resolution satellite imagery remains underexplored. This study investigates FSL methods for classifying Maasai bomas using PlanetScope imagery at 3.0– \(3.5~\text {m}\) ground sampling distance. To replicate operational constraints, we adopt an episodic evaluation framework where a small number of support samples per class and corresponding query samples are randomly selected from the dataset. We benchmark a state-of-the-art few-shot learning approach that integrate transfer learning and meta-learning strategies. For feature extraction, we evaluated Vision Transformers pretrained through DINO v1, DINO v2, and DINO v3, as well as convolutional architectures including ConvNeXt and Swin v2. Each backbone was evaluated with a Prototypical Network framework to measure its ability to generate discriminative embeddings under limited supervision. Furthermore, we compared models pretrained on out-of-domain datasets with those trained using in-domain satellite imagery to examine how pretraining domain influences generalization and adaptation in remote sensing tasks.