Abstract
Landform classification is a critical task for landscape evaluation, erosion studies, and hazard prediction. This study proposes an improved object-based classification framework for Chinese landform types using 1km Digital Elevation Model (DEM) data. By integrating Random Forest factor importance analysis with Gray-Level Co-occurrence Matrix (GLCM) texture analysis, the method automates the segmentation process and enhances classification accuracy. The approach achieved an overall accuracy of 71.4%, outperforming traditional unsupervised and object-based methods by 5.7% and 15.7%, respectively.
Introduction
Accurately categorizing the Earth's surface into landform types such as plains, hills, and mountains is fundamental for regional planning. Traditional pixel-based classification often suffers from "salt and pepper" noise—isolated misclassified pixels that disrupt spatial continuity. While Object-Based Image Analysis (OBIA) addresses this by grouping pixels into "objects," it typically relies on subjective, manual thresholding for segmentation. Furthermore, existing automated models often struggle with the high-dimensional complexity of terrain data. This research provides an objective, data-driven framework for large-scale landform classification across China's diverse topography.
Methodology
The proposed workflow transitions from manual parameter tuning to a quantitative, multi-stage process designed to reduce subjectivity and improve reliability.
Key Methodological Steps:
- Factor Selection & Redundancy Reduction: Nine terrain factors (e.g., slope, elevation, surface incision) were initially extracted. Sheffield’s entropy method and correlation analysis were employed to filter redundant data and identify the most informative combination of factors.
- Random Forest for Segmentation: To eliminate subjectivity in the segmentation phase, a Random Forest classifier was used to calculate the "importance" of each terrain factor. These importance values served as objective weights for multi-scale segmentation.
- Texture Analysis (GLCM): Recognizing that landforms possess distinct spatial patterns, the Gray-Level Co-occurrence Matrix (GLCM) was applied. This allowed the model to incorporate terrain "texture"—the spatial relationship between elevation values—rather than relying solely on raw height data.
- Feature Space Optimization (FSO): Euclidean distances between samples were calculated to determine the optimal feature combination that maximized the distinction between different landform classes.
Results and Discussion
The method was validated using a 1:1,000,000 scale DEM of China, with the 1:4,000,000 Chinese Geomorphological Map serving as the reference "ground truth."
Classification Performance
The proposed method achieved an overall accuracy of 71.4% with a Kappa coefficient of 0.67. This represents a significant improvement over standard ISODATA unsupervised classification and traditional object-based methods.
Class-Specific Insights
The model demonstrated high reliability in identifying "low-mid mountains" (84.1% user accuracy) and "plains" (83.7%). However, "low mountains" proved more challenging (47.7% accuracy), often being confused with hills or higher mountain ranges due to overlapping topographic signatures. Texture analysis via GLCM was particularly effective for "extreme high mountains," capturing their distinctively "broken" spatial patches more accurately than simple slope analysis.
Conclusion
By automating the "art" of segmentation through Random Forest weighting and incorporating spatial texture via GLCM, this study provides a robust framework for macro-scale landform mapping. While the method significantly improves accuracy, future work will focus on adaptive scaling for local terrain complexity and optimizing computational efficiency for higher-resolution datasets, such as 90m SRTM data.
Summary updated