Esrgan-Based Dem Super-Resolution for Enhanced Slope Deformation Monitoring in Lantau Island of Hong Kong

Wu, Zherong; Ma, Peifeng

doi:10.5194/isprs-archives-xliii-b3-2020-351-2020

Cited by 8 publications

(7 citation statements)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Demiray, et al [10] proposed D-SRGAN to increase the resolution up to four times based on the SRGAN method. Wu and Ma [9] introduced ESRGAN to specifically address the limited DEM resolution of landslide areas, which achieved tangible results and validated the practicability of the ESRGAN method on DEM SR. Notwithstanding, all these models generated unrealistic LR DEMs synthesized by simple degradation models (e.g., bicubic downsampling) to obtain the training data (LR DEMs with corresponding HR DEMs).…”

Section: B Dem Srmentioning

confidence: 93%

See 1 more Smart Citation

Real-World DEM Super-Resolution Based on Generative Adversarial Networks for Improving InSAR Topographic Phase Simulation

Wu¹,

Zhao²,

Ma³

et al. 2021

IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing

View full text Add to dashboard Cite

Topographic phase simulation is important for deformation estimation in differential synthetic aperture radar (SAR) interferometry (DInSAR). The most commonly used 30-m resolution SRTM digital elevation model (DEM) is usually required to be resampled due to its relatively low resolution (LR) comparing to the high resolution (HR) SAR images. Although the WorldDEM TM with a 12-m resolution achieves global coverage, it is not available freely. Consequently, it is useful to evaluate the practicability of the super-resolution (SR) from LR SRTM DEMs to HR WorldDEM TM ones, which has not been investigated. Most existing DEM SR models are trained with synthetic datasets in which the LR DEMs are downsampled from their HR counterparts. However, these models become less effective when applied to real-world scenarios due to the domain gap between the synthetic and real LR DEMs. In this paper, we constructed a realworld DEM SR dataset where the LR and HR DEMs were collected from SRTM and WorldDEM TM , respectively. An ESRGAN model was adapted to train on the dataset. Considering that the real LR-HR pairs may suffer from misalignment, we introduced the perceptual loss for better optimizing the model. Moreover, a logarithmic normalization was proposed to compress the wide elevation range and adjust the uneven distribution.We also pretrained the model using natural images since collecting sufficient HR DEMs is costly. Experiments demonstrate that the proposed method achieves near 0.69dB improvement of peak signal-to-noise ratio (PSNR). In addition, our method is also validated to improve the topographic phase simulation by 23.42% of MSE.

show abstract

Section: B Dem Srmentioning

confidence: 93%

“…Deep learning (DL) models have shown promising results on both image SR and DEM SR [7][8][9][10][11]. The generative adversarial network (GAN) models achieved the state-of-art performance for image SR [12,13].…”

mentioning

confidence: 99%

Real-World DEM Super-Resolution Based on Generative Adversarial Networks for Improving InSAR Topographic Phase Simulation

Wu¹,

Zhao²,

Ma³

et al. 2021

IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing

View full text Add to dashboard Cite

show abstract

“…Testing the model at the same location as it is trained also does not provide full information on the applicability of such methods to improve the quality of the freely available global dataset. Some researchers also used Shuttle Radar Topography Mission (SRTM) data which is available freely for super-resolution applications (Jiao et al, 2020;Wu & Ma, 2020), but in these cases, super sampling was done at very low resolution, producing a final resolution of 30 by 30 meters, which is too low for most/ many geoscience modelling studies.…”

Section: Introductionmentioning

confidence: 99%

Deep Learning-Based Super-Resolution of Digital Elevation Models in Data Poor Regions.

Dahal¹,

Bout²,

Westen³

et al. 2022

Preprint

View full text Add to dashboard Cite

In order to develop reliable models, the geoscientific community requires high-resolution data sets. However, the collection of such data is a persistent challenge due to the limitations of resources. The concept of super-resolution, a method from the field of machine learning, can be used to predict a high-resolution version of a low-resolution dataset to improve usability in geoscientific applications. However, thus far, super-resolution is predominantly used in image data with few cases on improving the scientific data but focusing on improving quality of same downsampled data. More importantly, it is unknown whether models that are developed and trained with high-resolution data of specific locations can also be applied to data-poor regions. To address these gaps, this study investigated the use of deep learning-based super-resolution to improve the resolution of digital elevation data, focusing on the question whether models trained with high resolution data can also be applied to regions for which only low-resolution data are available. We focused on Digital Elevation Models (DEMs), as these are among the most important datasets for many geoscientific applications and used two of the most advanced Super-Resolution models (EBRN and ESRGAN) from different groups of deep learning architecture. We trained those models extensively using high-resolution LiDAR DEM data from Austria, and found that, for the Austrian study sites, these models performed better than commonly used interpolation techniques such as bicubic interpolation. To test model applicability to different terrain conditions, we applied the models developed and trained with Austrian data to globally available free datasets on/for Colombia and Dominica. A novel loss function, training technique and evaluation metrics were developed to train and evaluate the results focusing on improving DEM data. Our results show that super-resolution can improve the accuracy of global datasets by 30-50% relative to bicubic interpolation, thus providing a promising solution for locations for which only low-resolution DEM data are available.

show abstract

“…Several others have used CNNs to reconstruct SR DEMs [36][37][38]. Some of the classical models in image SR such as EDSR, SRGAN, and ESRGAN, which have achieved good results, have been used in SR DEM applications [39][40][41].…”

Section: Introductionmentioning

confidence: 99%

An Enhanced Double-Filter Deep Residual Neural Network for Generating Super Resolution DEMs

et al. 2021

View full text Add to dashboard Cite

High-resolution DEMs are important spatial data, and are used in a wide range of analyses and applications. However, the high cost to obtain high-resolution DEM data over a large area through sensors with higher precision poses a challenge for many geographic analysis applications. Inspired by the convolution neural network (CNN) excellent performance in super-resolution (SR) image analysis, this paper investigates the use of deep residual neural networks and low-resolution DEMs to generate high-resolution DEMs. An enhanced double-filter deep residual neural network (EDEM-SR) method is proposed, which uses filters with different receptive field sizes to fuse and extract features and reconstruct a more realistic high-resolution DEM. The results were compared with those generated with the bicubic, bilinear, and EDSR methods. The numerical accuracy and terrain feature preserving effects of the EDEM-SR method can generate reconstructed DEMs that better match the original DEMs, show lower MAE and RMSE, and improve the accuracy of the terrain parameters. MAE is reduced by about 30 to 50% compared with traditional interpolation methods. The results show how the EDEM-SR method can generate high-resolution DEMs using low-resolution DEMs.

show abstract

Esrgan-Based Dem Super-Resolution for Enhanced Slope Deformation Monitoring in Lantau Island of Hong Kong

Cited by 8 publications

References 13 publications

Real-World DEM Super-Resolution Based on Generative Adversarial Networks for Improving InSAR Topographic Phase Simulation

Real-World DEM Super-Resolution Based on Generative Adversarial Networks for Improving InSAR Topographic Phase Simulation

Deep Learning-Based Super-Resolution of Digital Elevation Models in Data Poor Regions.

An Enhanced Double-Filter Deep Residual Neural Network for Generating Super Resolution DEMs

Contact Info

Product

Resources

About