当前位置: X-MOL 学术Remote Sens. Environ. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Integrating topographic knowledge into deep learning for the void-filling of digital elevation models
Remote Sensing of Environment ( IF 13.5 ) Pub Date : 2021-12-02 , DOI: 10.1016/j.rse.2021.112818
Sijin Li 1, 2, 3, 4 , Guanghui Hu 1, 2, 3 , Xinghua Cheng 5 , Liyang Xiong 1, 2, 3 , Guoan Tang 1, 2, 3 , Josef Strobl 2, 4
Affiliation  

Digital elevation models (DEMs) contain some of the most important data for providing terrain information and supporting environmental analyses. However, the applications of DEMs are significantly limited by data voids, which are commonly found in regions with rugged terrain. We propose a novel deep learning-based strategy called a topographic knowledge-constrained conditional generative adversarial network (TKCGAN) to fill data voids in DEMs. Shuttle Radar Topography Mission (SRTM) data with spatial resolutions of 3 and 1 arc-seconds are used in experiments to demonstrate the applicability of the TKCGAN. Qualitative topographic knowledge of valleys and ridges is transformed into new loss functions that can be applied in deep learning-based algorithms and constrain the training process. The results show that the TKCGAN outperforms other common methods in filling voids and improves the elevation and surface slope accuracy of the reconstruction results. The performance of the TKCGAN is stable in the test areas and reduces the error in the regions with medium and high surface slopes. Furthermore, the analysis of profiles indicates that the TKCGAN achieves better performance according to a visual inspection and quantitative comparison. In addition, the proposed strategy can be applied to DEMs with different resolutions. This work is an endeavour to transform topographic knowledge into computer-processable rules and benefits future research related to terrain reconstruction and modelling.



中文翻译:

将地形知识集成到深度学习中以填充数字高程模型

数字高程模型 (DEM) 包含一些最重要的数据,用于提供地形信息和支持环境分析。然而,DEM 的应用受到数据空白的极大限制,数据空白通常出现在地形崎岖的地区。我们提出了一种新的基于深度学习的策略,称为地形知​​识约束条件生成对抗网络 (TKCGAN),以填补 DEM 中的数据空白。在实验中使用空间分辨率为 3 和 1 弧秒的航天飞机雷达地形任务 (SRTM) 数据来证明 TKCGAN 的适用性。山谷和山脊的定性地形知识转化为新的损失函数,可应用于基于深度学习的算法并限制训练过程。结果表明,TKCGAN在填充空隙方面优于其他常用方法,提高了重建结果的高程和地表坡度精度。TKCGAN 在测试区域的性能稳定,在中高地表坡度区域减少了误差。此外,轮廓分析表明,根据目视检查和定量比较,TKCGAN 实现了更好的性能。此外,所提出的策略可以应用于具有不同分辨率的 DEM。这项工作致力于将地形知识转化为计算机可处理的规则,并有利于未来与地形重建和建模相关的研究。TKCGAN 在测试区域的性能稳定,在中高地表坡度区域减少了误差。此外,轮廓分析表明,根据目视检查和定量比较,TKCGAN 实现了更好的性能。此外,所提出的策略可以应用于具有不同分辨率的 DEM。这项工作致力于将地形知识转化为计算机可处理的规则,并有利于未来与地形重建和建模相关的研究。TKCGAN 在测试区域的性能稳定,在中高地表坡度区域减少了误差。此外,轮廓分析表明,根据目视检查和定量比较,TKCGAN 实现了更好的性能。此外,所提出的策略可以应用于具有不同分辨率的 DEM。这项工作致力于将地形知识转化为计算机可处理的规则,并有利于未来与地形重建和建模相关的研究。所提出的策略可以应用于具有不同分辨率的 DEM。这项工作致力于将地形知识转化为计算机可处理的规则,并有利于未来与地形重建和建模相关的研究。所提出的策略可以应用于具有不同分辨率的 DEM。这项工作致力于将地形知识转化为计算机可处理的规则,并有利于未来与地形重建和建模相关的研究。

更新日期:2021-12-02
down
wechat
bug