Object-based large-scale terrain classification combined with segmentation optimization and terrain features: A case study in China

Terrain classification involves essential tasks in geomorphology, landscape investigation, regional planning, and hazard prediction. Most existing methods are based on a simple thresholding approach. However, such an approach is limited in terms of accuracy and robustness, especially for large-scale tasks. To overcome this limitation, this article proposes an object-based framework combined with the random forest. Six terrain factors, namely terrain relief, surface roughness, elevation, elevation coefficient variation, shaded relief, and accumulative curvature, are first selected by correlation analysis using Sheffield's entropy. The obtained segmentation result is then optimized by Moran's I and the weighted variance, combining both terrain factors and textures derived from digital elevation models. Then, the features are selected among both terrain factors and their gray-level co-occurrence matrix textures. Finally, the features are fed into the random forest classifier. Seven landform types are classified, including plain, hill, low mountain, low-middle mountain, high-middle mountain, high mountain, and extremely high mountain. A case study in China was conducted and achieved an overall accuracy of 80.53% compared with the official landform atlas, which is better performance over the compared semi-automatic methods. The transferability of our framework was further confirmed by an additional application in provincial-scale mapping with a different classification system.

Automated summary

This study proposed an object-based framework combined with random forest for terrain classification, improving accuracy by selecting terrain factors and textures. In a case study in China, an overall accuracy of 80.53% was achieved compared to the official landform atlas, outperforming other semi-automatic methods.