TL;DR
This paper introduces a novel "regional fusion" (RF) algorithm that combines slope variation analysis with an optimized region-growing technique to accurately extract loess shoulder-lines from Digital Elevation Models (DEMs). This method significantly improves upon existing algorithms, particularly in challenging areas with gentle slopes, achieving high accuracy in identifying these crucial geomorphological features for erosion assessment and landform classification.
What Problem Does This Paper Address?
The Loess Plateau in China is characterized by unique positive and negative terrains (P-N terrains) that are critical for understanding soil erosion and landform evolution. Loess shoulder-lines serve as the primary boundary between these P-N terrains and are vital for identifying gully heads and predicting erosion. However, existing methods for automatically extracting these shoulder-lines, especially in areas with subtle slope changes, often struggle with accuracy and continuity. Manual extraction is time-consuming and inefficient, necessitating the development of more robust automated approaches.
Methods
The proposed "regional fusion" (RF) method integrates two main components:
- Initial Shoulder-Line and Negative Region Identification:
- A slope-variation-based method is used to identify initial candidate points for shoulder-lines. This involves analyzing the difference between smoothed and real DEMs to delineate the negative terrain.
- A region-growing algorithm, with optimized growth standards incorporating topographic factors like slope and Topographic Position Index (TPI), is employed to expand these initial points and generate candidate regions for both shoulder-lines and negative terrains.
- Fusion and Refinement:
- The identified candidate shoulder-line regions and negative regions are fused to create a continuous area.
- Noise and terrace artifacts are eliminated using an octree search algorithm.
- A 3x3-window edge detector is applied to extract the shoulder-line contour.
- Finally, a hit-or-miss transform (HMT) is used to remove spurious points (spurs and congregate points) and optimize the extracted shoulder-lines.
The core innovation lies in the optimized region-growing algorithm, which uses a weighted combination of topographic features (slope and TPI) to define its growth standards, making it more adaptable to complex terrain.
Datasets
- Digital Elevation Models (DEMs): 5-meter resolution DEMs were generated from fieldwork using DJI drones with 60% overlap, processed with Photoscan software. Data was collected for two study areas in the Loess Plateau: Suide and Yan'an.
- Orthophoto Model (DOM): 0.1-meter resolution DOM data was used for manual interpretation to create a reference dataset for accuracy assessment.
Key Insights and Ideas
- Region-Growing Optimization is Key: The paper demonstrates that simply applying standard region-growing isn't enough. By incorporating multiple topographic features (slope, TPI) into the growth criteria, the algorithm becomes much more sensitive to subtle terrain changes, crucial for accurate shoulder-line extraction.
- Fusion for Continuity: Merging the initially identified shoulder-line and negative regions is essential for creating a continuous line, overcoming the inherent fragmentation of initial detection methods.
- Multi-Stage Refinement: The combination of edge detection and morphological operations (HMT) is vital for cleaning up the extracted lines, removing artifacts that can arise from automated processes.
- Slope Variation as a Strong Indicator: The study highlights the importance of slope variation, particularly the difference between upstream and downstream slopes, in identifying potential shoulder-line points.
Key Results
- High Accuracy: The proposed RF method achieved a Euclidean distance offset percentage (EDOP) accuracy of 96.9% within a 10-meter deviation range compared to manually digitized shoulder-lines.
- Superior Performance: When compared to Zhou's method, the improved snake model (ISM), and the bidirectional DEM relief-shading (BRS) method, the RF method consistently outperformed them in terms of EDOP, Mean Absolute Error (MAE), and Standard Absolute Deviation (SAD).
- Robustness: The RF method demonstrated superior performance in extracting shoulder-lines in areas with gentle slopes and complex terrains, where other methods struggled with discontinuity or roughness.
- Parameter Sensitivity: The study found that the highest accuracy was achieved with specific weights for slope and TPI (α=0.7, β=0.3), suggesting slope plays a more dominant role in the growth standard. However, the overall precision was not overly sensitive to these parameters, with a=0.5 and β=0.5 also yielding good results.
Limitations and Future Work
- Uncertainties near Ridges: The paper acknowledges that some uncertainties remain in regions close to ridgelines, suggesting further refinement is needed.
- Limited Terrain Factors: The current method uses only two topographic factors (slope and TPI) for the growing standard. Future work could explore incorporating more factors for enhanced robustness.
- Generalizability: The study focused on loess ridges and hills. Testing the method's generalizability on loess tableland landforms is recommended.
- DEM Resolution: The study utilized 5-meter resolution DEMs. Investigating the impact of different DEM resolutions on extraction accuracy is a potential future direction.
Summary updated