Super resolution aided multi hazard modeling: Is it possible?

Kurzfassung

The losses due to natural hazards are very high and show an increasing trend due to climate change; human and economic growth; and unplanned development. The risk due to those hazards can be reduced using multi-hazard risk assessment using hazard, the element at risk and vulnerability data. However, due to the lack of good quality and high-resolution data in developing nations, modelling hazards at infrastructure level is difficult. Deep learning-based Super-Resolution can be a solution to increase the spatial resolution of freely available global datasets. However, no studies exist that produced high-resolution output from globally available DEM data using Super-Resolution to improve the quality of physically based hazard modelling. Furthermore, due to differences in data collection sources and value ranges in DEMs, they cannot be compared in absolute values, and there is a lack of techniques to evaluate the improvement done with Super-Resolution in geoscientific data. Moreover, none of the existing research has trained the Super-Resolution models in one region and applied them to another region. To address these problems, our research aimed to increase the applicability of physically based models in data-poor regions by improving the spatial resolution of globally available datasets by using deep learning-based Super-Resolution. To fulfil our objectives, we selected to work on Digital Elevation Models as the target variable due to its importance in hazard modelling and global availability. We used the two of the most advanced Super-Resolution models (EBRN and ESRGAN), each from different groups of deep learning architecture. These models were trained extensively using high-resolution LiDAR DEM data from Austria. After proving that they perform better than most used interpolation techniques such as bicubic interpolation in the study areas in Austria, they were applied in globally available free datasets in Colombia and Dominica. Furthermore, novel loss function and evaluation metrics were developed to train and evaluate the results focusing on improving DEM data. Furthermore, physically based modelling was used to evaluate the impact of Super-Resolution in multi-hazard modelling. We used 21 different scenarios to test the applicability of Super-Resolution compared to existing interpolation techniques and global commercial data. Each scenario was calibrated for 20 iterations (total 420 iterations, ~5460 CPU hours) in Microsoft Azure, which is the first time that OpenLISEM was used in a cloud computing environment. The results were evaluated in terms of the modelled extent of hazardous processes, the height of flow, and the time of solid and fluid flow to prove the applicability of the Super-Resolution approach. The analysis shows that the use of global DEM data with Super-Resolution processing was able to increase the accuracy of hazard modelling output as compared to DEMs made with existing interpolation techniques. Furthermore, when evaluating derivative DEM products through visual analysis, it is observed that the Super-Resolution has increased the crispness of valley lines and ridgelines in the DEM datasets. However, the specific topographic features that are not present in low-resolution data could not be reconstructed using the Super-Resolution, limiting its use in geomorphological mapping. The applicability of Super-Resolution was tested in multiple locations, and it could prove that the technique resulted in 8-25% improvement in all of the study sites. The results also show that the capacity of both models (EBRN and ESRGAN) is generally very similar. There are a few challenges in calibration, such as the use of Gradient Descent requiring more iterations and the lack of datasets and metrics to compare our results with existing Super-Resolution models. Furthermore, we could not compare our results on multi-hazard modelling to other research because there is no published work using the Super-Resolution in multi-hazard modelling.