Scholarlink

Satellite image classification is an essential task in remote sensing with applications in urban development, agriculture, and environmental monitoring. Although deep learning has improved classification accuracy, training efficiency remains a challenge. This study investigated progressive resolution training as a coarse-to-fine alternative to direct high-resolution training using EfficientNet B3, MobileNetV3, and ConvNeXt on the UC Merced Land Use and SIRI WHU datasets. The results show that ConvNeXt with Bilinear interpolation achieved 98.41% accuracy on UC Merced. This outperformed direct training (97.93%). It also reduced the training time by 34.5% (from 1 h 46 m 42 s to 1 h 11 m 19 s). On SIRI WHU, the best performance reached 98.33% with up to 21.5% faster training. These findings suggest that progressive resolution training enhances both accuracy and efficiency. This makes it a promising strategy for scalable satellite image classification.

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

Learning Coarse-to-Fine: Progressive Resolution Training Strategies for Efficient Transfer Learning in Satellite Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Quang Nhat Nguyen,
  • Nguyen Giap Dang,
  • Cao Vu Bui

摘要

Satellite image classification is an essential task in remote sensing with applications in urban development, agriculture, and environmental monitoring. Although deep learning has improved classification accuracy, training efficiency remains a challenge. This study investigated progressive resolution training as a coarse-to-fine alternative to direct high-resolution training using EfficientNet B3, MobileNetV3, and ConvNeXt on the UC Merced Land Use and SIRI WHU datasets. The results show that ConvNeXt with Bilinear interpolation achieved 98.41% accuracy on UC Merced. This outperformed direct training (97.93%). It also reduced the training time by 34.5% (from 1 h 46 m 42 s to 1 h 11 m 19 s). On SIRI WHU, the best performance reached 98.33% with up to 21.5% faster training. These findings suggest that progressive resolution training enhances both accuracy and efficiency. This makes it a promising strategy for scalable satellite image classification.