1:45 PM - 2:00 PM
[HTT17-01] Terrain Analysis of Handayama Using DEM Modeling by Machine Learning
Keywords:Digital Elevation Model, Modeling, Machine Learning, Kriging, Slope Analysis
In recent years, the digitization of terrain, known as Digital Elevation Models (DEMs), has advanced, with an increasing trend in providing them as open data. Furthermore, these DEMs have become highly resolutioned, allowing access to open data with a pitch as fine as 1 meter. However, as these high-resolution DEMs include data on features such as large boulders at a scale of a few meters, modeling the overall terrain, such as hills and mountains, requires the use of Envelope Plane modeling, such as Summit Level Maps. This presentation report on the application of machine learning to DEM modeling.
The target terrain for modeling includes Handayama (elevation 151m) and Tatsunokuchiyama (elevation 257m) located in Okayama City. The digital elevation models utilized are based on the 10m mesh (DEM10b) from the National Institute of Land and Infrastructure Management (Geospatial Information Authority of Japan, 2024). The modeling process involves using GNU R (R Core Team, 2024), the machine learning package kernlab (Karatzoglou, Smola, and Hornik, 2023), and the geostatistics package gstat (Gräler, Pebesma, and Heuvelink, 2016). QGIS (QGIS Development Team, 2023) is used for map representation, and the Contour plugin (QGIS Development Team, 2023) is employed for contour creation.
Initially, the spatial structure of DEM data for Handayama and Tatsunokuchiyama was estimated using variograms, and an interim terrain model (OK model) was obtained through ordinary kriging. Subsequently, the machine learning algorithm was fine-tuned using the OK model, and the DEM10b was trained to infer a 20m pitch terrain model (ML model). The comparison between DEM and ML models is shown in Fig.1. Further, slope analysis of the ML model was conducted, and a slope histogram was generated. Lastly, peak separation in the slope histogram was performed to investigate the characteristics of slope distribution. As a result, it was observed that Handayama and Tatsunokuchiyama, where continuous geological formations are presumed, exhibit different slope distribution characteristics. This is attributed to geological activity, possibly related to the distribution of rhyolite not found in Handayama, in the northeast of Tatsunokuchiyama.
The OK model of the DEM ensures that the envelope surface always passes through the original points of the DEM. On the other hand, depending on the tuning of the algorithm, the ML model may infer an overfitting model that perfectly reproduces the DEM. Therefore, both models are challenging to use when considering the overall terrain, such as hills and mountains. However, it is believed that a suitable model for terrain analysis can be obtained by using both methods in combination.
The target terrain for modeling includes Handayama (elevation 151m) and Tatsunokuchiyama (elevation 257m) located in Okayama City. The digital elevation models utilized are based on the 10m mesh (DEM10b) from the National Institute of Land and Infrastructure Management (Geospatial Information Authority of Japan, 2024). The modeling process involves using GNU R (R Core Team, 2024), the machine learning package kernlab (Karatzoglou, Smola, and Hornik, 2023), and the geostatistics package gstat (Gräler, Pebesma, and Heuvelink, 2016). QGIS (QGIS Development Team, 2023) is used for map representation, and the Contour plugin (QGIS Development Team, 2023) is employed for contour creation.
Initially, the spatial structure of DEM data for Handayama and Tatsunokuchiyama was estimated using variograms, and an interim terrain model (OK model) was obtained through ordinary kriging. Subsequently, the machine learning algorithm was fine-tuned using the OK model, and the DEM10b was trained to infer a 20m pitch terrain model (ML model). The comparison between DEM and ML models is shown in Fig.1. Further, slope analysis of the ML model was conducted, and a slope histogram was generated. Lastly, peak separation in the slope histogram was performed to investigate the characteristics of slope distribution. As a result, it was observed that Handayama and Tatsunokuchiyama, where continuous geological formations are presumed, exhibit different slope distribution characteristics. This is attributed to geological activity, possibly related to the distribution of rhyolite not found in Handayama, in the northeast of Tatsunokuchiyama.
The OK model of the DEM ensures that the envelope surface always passes through the original points of the DEM. On the other hand, depending on the tuning of the algorithm, the ML model may infer an overfitting model that perfectly reproduces the DEM. Therefore, both models are challenging to use when considering the overall terrain, such as hills and mountains. However, it is believed that a suitable model for terrain analysis can be obtained by using both methods in combination.