Scholarlink

<p>With more and more remote sensing data available on a&#xa0;global scale, the Earth observation community strives to harness the power of modern deep learning techniques by developing globally applicable models. However, remote sensing images exhibit strongly heterogeneous, geolocation-dependent characteristics, making this a&#xa0;challenging endeavor. In this paper, we introduce the geolocation-aware deep coding strategy to incorporate geolocation information of remote sensing data into the training of the deep learning models. The proposed method consists of defining regional subnetworks dedicated to each subset of the dataset with similar geolocational characteristics. Using two application examples, namely the mapping of building footprints from multi-spectral Sentinel‑2 imagery, and the task of forest detection from single-channel thermal infrared Landsat imagery, we show that the proposed deep coding strategy stabilizes the training performance and can also improve the predictive power of deep neural networks designed for remote sensing data analysis.</p>

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

Geolocation-Aware Deep Coding

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年1月7日

查看全文
  • Mojgan Madadikhaljan,
  • Michael Schmitt

摘要

With more and more remote sensing data available on a global scale, the Earth observation community strives to harness the power of modern deep learning techniques by developing globally applicable models. However, remote sensing images exhibit strongly heterogeneous, geolocation-dependent characteristics, making this a challenging endeavor. In this paper, we introduce the geolocation-aware deep coding strategy to incorporate geolocation information of remote sensing data into the training of the deep learning models. The proposed method consists of defining regional subnetworks dedicated to each subset of the dataset with similar geolocational characteristics. Using two application examples, namely the mapping of building footprints from multi-spectral Sentinel‑2 imagery, and the task of forest detection from single-channel thermal infrared Landsat imagery, we show that the proposed deep coding strategy stabilizes the training performance and can also improve the predictive power of deep neural networks designed for remote sensing data analysis.