Geodaysit 2023

Authors: A. Capolupo & E. Tarantino
Several research involving Earth's physical processes and depicting environmental systems are computationally time-consuming, and as a result, have a substantial impact on the time necessary to collect and manage the data. Over the years, numerous acceptable methods for describing surface morphology and enabling quick computer solutions were developed. Nevertheless, since 1991, Digital Elevation Model (DEM) has been recognized as the finest alternative for attaining this goal because, in addition to its capacity to provide baseline morphological information quickly, it also has the exclusive property of being a 2.5-D surface. The quality and trustworthiness of the results provided by its use are determined by its resolution, elevation accuracy, and shape/topological correctness. Elevation accuracy is normally established by statistically analysing differences between DEMs and reference datasets such as Ground Control Points (GCPs), whereas shape/topological correctness is typically defined by demonstrating DEM conformity with some universal principles. Therefore, the root mean square error is commonly used to achieve the first aim, whilst DEM derivates are examined in the second one. However, neither approach is without limits since their performance is influenced by the quality of the reference data and the complexity in measuring DEM realism.
This is much more difficult when the DEM under consideration encompasses the entire globe. Even though they are described as a homogenous product, the accuracy of Global DEMs in terms of elevation and realism varies according to geographical location and morphology, land cover, and climate. Furthermore, as satellite stereoscopic technologies, as well as photogrammetric and SAR interferometric methods, have evolved, the amount of Global DEMs collected has substantially increased. Most of them were also collected in different historical periods and, consequently, they may be useful free open-source data for conducting a consistent global study change detection analysis.
In such a framework, this study aims to investigate the appropriateness of medium-resolution open-access Global DEMs in evaluating changes in urban contexts between 2000 and 2011. To accomplish this, the primary freely accessible Global DEMs were statistically examined, and after selecting the best pair, a change detection analysis was carried out. To assess its accuracy, the findings were compared to the Copernicus Land Monitoring service's land use layers from the same historical periods (https://land.copernicus.eu/). Lastly, this study seeks to estimate and predict the caused by building density bias in accordance with the urban fabric type.
The procedure was implemented by writing appropriate Java-script code on the Google Earth Engine (GEE) web-based platform. Hence, the GEE catalogue was first consulted to determine the available Global DEMs corresponding to the historical period under investigation, and, once identified, they were imported into the application programming interface and validated using the "internal" technique. As a result, AW3D30 (3.2), which was launched in early January 2021, and SRTM DEM V3 were deemed the optimal combination for research purposes during an 11-year timeframe. Thus, they were used as input data for calculating the corresponding DEM of Differences (DoD) and quantify the alteration in urban environments. Owing to the law propagation error, the resultant DoD had substantial internal incoherencies, which were subsequently statistically eliminated by using the Tukeys' filter. This is widely acknowledged as an effective method for identifying and cleaning out internal noise without prior awareness of it. Yet, a significant amount of Tukey's outliers was identified and eliminated in their respective DoD, mostly in wooded and hilly zones, owing to differing degrees of quality of the input data. Following that, to reduce misclassification and distinguish noise from real changes, the resulting DoD was further filtered using the Uniformly Distributed Error (UDE) strategy, developed by Brasington et al. in 2003. However, the UDE technique, while exploiting a gaussian distribution of internal error, does not adapt the filtering threshold to the local conditions, resulting in an over or underestimation of the amount of information to remove. Urban variation was now assessed by combining the filtered DoD result with Corine Land Cover (CLC) data. This integration also enabled statistical investigation and modelling of the DoD error associated with urban fabric type. When comparing the CLC information to both Tukey's outliers and UDE noise in urban areas, it is discovered that error increased linearly with building density. This implies that urban changes quantification could be improved further by correcting the building density bias. In future works, the introduced approach will be enhanced by taking building height into consideration.