Deep learning-based automated terrain classification using high-resolution DEM data

Landforms are a fundamental component of the natural environment, and digital terrain mapping on a large spatial scale is important when studying landforms. In this study, we adopted a semantic segmentation model in computer vision to classify elementary landform types using AW3D30 digital elevation model (DEM) data. We built a semantic segmentation model with an FCN-ResNet architecture that extracts features using a residual network (ResNet) and obtains pixel-level segmentation of the DEM using a fully convolutional network (FCN). A lightweight decoder based on skip connections was adopted to maintain detailed information at different scales. We used the 1:1,000,000 Chinese landform map as the label and tested different combinations of terrain factors. The experiments indicate that increasing the terrain factors has no significant influence on the model, and the semantic information can be learned using only DEM data. The model has strong feature extraction capability and is tolerant to noise and error. The results of landform category prediction confirm that deep learning methods have strong potential for landform classification and will have great application prospects in the field of geomorphological research.

Automated summary

In this study, a semantic segmentation model was used to classify elementary landform types, and digital terrain mapping using AW3D30 DEM data was crucial in studying landforms. A semantic segmentation model with an FCN-ResNet architecture was built to extract features using a residual network (ResNet) and achieve pixel-level segmentation of the DEM. Results showed that increasing terrain factors had no significant impact on the model, and semantic information could be learned solely from DEM data. The model demonstrated strong feature extraction capability and tolerance to noise and error, and deep learning methods showed great potential in landform classification for geomorphological research.