针对现有数字高程模型（digital elevation model,DEM）数据空洞填充算法存在修复效果不连续、适用空值范围狭小以及细节重构丢失等问题，提出了一种融合自注意力机制的生成式对抗网络的DEM空洞填充方法。首先，构造自注意力机制提取DEM数据特征信息，改善DEM空洞填充结果高程值不连续和纹理细节缺失的问题。其次，在生成器中使用对称结构的卷积与反卷积网络结构，保证生成可靠度较高的数据以实现空洞区域填充，并利用判别器实现空洞填充结果的预分类。最后，结合重构损失函数进行训练，提升DEM空洞填充结果对异常值的鲁棒性，增强模型的回归能力。采用不同分辨率DEM数据进行空洞填充并与现有方法进行对比，结果表明：所提方法能够大幅提升填充精度，有效解决原始数据中存在的空洞问题。
Aiming at the problems of existing DEM data void filling algorithms, such as discontinuous repair effect, narrow applicable null value range and loss of detail reconstruction, this paper proposes a DEM void filling method integrating self-attention mechanism with generative adjunctive network. Firstly, a self-attention mechanism is constructed to extract DEM data feature information to improve the elevation discontinuity and texture detail loss of DEM cavity filling results. Secondly, symmetric convolutional and deconvolution network structures are used in the generator to ensure the generation of high reliability data to realize the filling of the void region, and the discriminator is used to realize the pre-classification of the filling results. Finally, combined with the reconstruction of loss function and the generation of adversarial loss function, the network training was carried out to improve the robustness of DEM cavity filling results to outliers and enhance the regression ability of the model. The experimental results show that compared with the filling results of spatial interpolation and deep learning, the proposed method can greatly improve the filling accuracy and effectively solve the problems of holes in the original data.
[1] Li D R. Towards the development of remote sensing and GIS in the 21st century[J]. Geomatics and Information Science of Wuhan University, 2003, 28(2):127-131.
[2] Jafarzadegan K, Merwade V. A DEM-based approach for large-scale floodplain mapping in ungauged watersheds[J]. Journal of Hydrology, 2017, 550:650-662.
[3] Toutin T. Impact of terrain slope and aspect on radargrammetric DEM accuracy[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2002, 57(3):228-240.
[4] Polat N, Uysal M. An experimental analysis of digital elevation models generated with Lidar data and UAV photogrammetry[J]. Journal of the Indian Society of Remote Sensing, 2018, 46(7):1135-1142.
[5] 李德仁, 周月琴, 马洪超. 卫星雷达干涉测量原理与应用[J]. 测绘科学, 2000, 25(1):9-12. Li D R, Zhou Y Q, Ma H C. Principle and application of satellite radar interferometry[J]. Science of Surveying and Mapping, 2000, 25(1):9-12. (in Chinese)
[6] Gonga-Saholiariliva N, Gunnell Y, Petit C, et al. Techniques for quantifying the accuracy of gridded elevation models and for mapping uncertainty in digital terrain analysis[J]. Progress in Physical Geography, 2011, 35(6):739-764.
[7] Reuter H I, Nelson A, Jarvis A. An evaluation of void-filling interpolation methods for SRTM data[J]. International Journal of Geographical Information Science, 2007, 21(9):938-1008.
[8] Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative adversarial networks[J]. Communications of the ACM, 2020, 63(11):139-144.
[9] Dong G, Chen F, Ren P. Filling SRTM void data via conditional adversarial networks[C]//2018 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2018:7441-7443.
[10] Di Z, Ximeng C, Fan Z, et al. Spatial interpolation using conditional generative adversarial neural networks[J]. International Journal of Geographical Information Science, 2020, 34(4):735-758.
[11] Gavriil K, Muntingh G, Barrowclough O. Void filling of digital elevation models with deep generative models[J]. IEEE Geoscience and Remote Sensing Letters, 2019, 16(10):1645-1649.
[12] Zhang C, Shi S, Ge Y, et al. DEM void filling based on context attention generation model[J]. ISPRS International Journal of Geo-information, 2020, 9(12):734.
[13] Qiu Z, Yue L, Liu X. Void filling of digital elevation models with a terrain texture learning model based on generative adversarial networks[J]. Remote Sensing, 2019, 11(23):2829.
[14] 赵露露, 沈玲, 洪日昌. 图像修复研究进展综述[J]. 计算机科学, 2021, 48(3):14-26. Zhao L L, Shen L, Hong R C. Survey on image inpainting research progress[J]. Computer Science, 2021, 48(3):14-26. (in Chinese)
[15] Zhang H, Goodfellow I, Metaxas D, et al. Self-attention generative adversarial networks[C]//International Conference on Machine Learning, 2019:7354-7363.
[16] 李海燕, 吴自莹, 郭磊, 等. 基于混合空洞卷积网络的多鉴别器图像修复[J]. 华中科技大学学报(自然科学版), 2021, 49(3):40-45. Li H Y, Wu Z Y, Guo L, et al. Multi-discriminator image repair based on hybrid dilated convolutional networks[J]. Journal of Huazhong University of Science and Technology (Natural Science Edition), 2021, 49(3):40-45. (in Chinese)
[17] Radford A, Metz L, Chintala S. Unsupervised representation learning with deep convolutional generative adversarial networks[DB/OL]. 2015[2021-10-10]. https://arxiv.org/abs/1511.06434.
[18] Mnih V, Heess N, Graves A, et al. Recurrent models of visual attention[C]//In Proceedings of the 27th International Conference on Neural Information Processing Systems, 2014, 2204-2212.
[19] Miyato T, Kataoka T, Koyama M, et al. Spectral normalization for generative adversarial networks[DB/OL]. 2018[2021-10-10]. https://arxiv.org/abs/1802.05957.
[20] Kingma D, Ba J. Adam:a method for stochastic optimization[DB/OL]. 2014[2021-10-10]. https://arxiv.org/abs/1412.6980.
