TL;DR
This paper introduces MFD-md, a novel multiple-flow-direction (MFD) algorithm that adaptively adjusts its flow partitioning based on local terrain steepness. This approach overcomes the limitations of fixed-exponent MFDs, leading to more accurate flow accumulation estimates, especially in complex terrain, and demonstrating lower error rates compared to existing methods.
What Problem Does This Paper Address?
Accurately modeling water flow across landscapes is crucial for various environmental applications, including hydrological modeling, soil erosion prediction, and geomorphological analysis. A key component of these models is determining how water flows from one point to another, often represented by "flow direction algorithms." Traditional methods fall into two categories: Single Flow Direction (SFD), which assigns all flow to a single downslope neighbor, and Multiple Flow Direction (MFD), which distributes flow among multiple downslope neighbors.
While MFD algorithms are generally better at capturing the spatial distribution of hydrological parameters, a critical challenge lies in how they partition flow. Most MFDs use a fixed "flow-partition exponent" (p) that dictates the distribution. This fixed value, however, fails to account for the fact that local terrain conditions (e.g., steepness, concavity) significantly influence how water disperses. A steep slope might lead to more concentrated flow, while a gentle slope might result in more diffuse flow. Applying a single exponent across an entire watershed, regardless of local terrain, leads to inaccuracies. This paper addresses this limitation by proposing an adaptive MFD approach that dynamically adjusts the flow-partition exponent based on local topographic attributes.
Methods
The core innovation of this paper is the development of an adaptive approach to determine the flow-partition exponent, denoted as 'p'. Instead of using a fixed value, MFD-md calculates 'p' dynamically for each cell based on a local topographic attribute.
The process involves two main steps:
- Selection of a Local Topographic Attribute: The authors explored several candidates, including maximum, minimum, and average downslope gradients. They determined that the maximum downslope gradient is the most suitable attribute because it best reflects the steepness of the steepest descent direction and is less sensitive to subtle DEM errors compared to other gradients.
- Construction of a Flow-Partition Function: A linear function, $f(e)$, was developed to map the maximum downslope gradient ($e$) to the flow-partition exponent ($p$). This function is designed such that a steeper gradient (larger $e$) results in a higher $p$ (approaching SFD behavior for convergent flow), while a gentler gradient (smaller $e$) results in a lower $p$ (promoting more dispersion, akin to MFD for divergent flow). The function is defined with specific lower ($p_l = 1.1$) and upper ($p_u = 10$) bounds for the exponent, representing near-complete dispersion and near-single flow direction, respectively. The function is parameterized with $e_{min}=0$ and $e_{max}=1$ (corresponding to a 45° slope), which are chosen to illustrate the adaptive capability.
The new MFD algorithm, named MFD-md, then uses this adaptive exponent in its flow partitioning equation:
$d_i = \frac{(\tan \beta_i)^{f(e)} \times L_i}{\sum_{j=1}^{8} (\tan \beta_j)^{f(e)} \times L_j}$
where $d_i$ is the fraction of flow to the $i$-th neighbor, $\tan \beta_i$ is the slope gradient, $L_i$ is the contour length factor, and $f(e)$ is the adaptive flow-partition function.
Datasets
The paper uses two types of datasets for evaluation:
- Artificial DEMs: Four mathematical surfaces were created to simulate typical terrain conditions: ellipsoid (convex slopes), inverse ellipsoid (concave slopes), saddle (ridges), and plane (straight slopes). These surfaces allowed for the calculation of "true" Specific Catchment Area (SCA) values, enabling quantitative error assessment.
- Real DEM: A Digital Elevation Model (DEM) of a small catchment in the Nenjiang watershed, Northeast China, was used for a practical application of the MFD-md algorithm. This DEM was pre-processed to fill depressions and amend flat areas.
Key Insights and Ideas
- Local Terrain Matters: The fundamental insight is that flow partitioning is not uniform across a landscape. Local terrain characteristics, particularly slope steepness, are the primary drivers of how water disperses.
- Adaptive Exponent is Key: A fixed flow-partition exponent is a significant oversimplification. An adaptive approach that responds to local terrain conditions is essential for accurate flow modeling.
- Maximum Downslope Gradient as a Proxy: The maximum downslope gradient effectively captures the local terrain's influence on flow direction and is robust enough to be a reliable indicator for adapting the flow-partition exponent.
- MFD-md's Versatility: The MFD-md algorithm demonstrates a remarkable ability to mimic SFD behavior in steep, convergent areas while behaving more like a traditional MFD in gentle, divergent areas, thus bridging the gap between these two approaches.
- Lower Complexity: Unlike some other adaptive MFD methods that rely on pre-computed flow accumulation, MFD-md's adaptive nature is based on local attributes, potentially leading to a simpler and more independent algorithm.
Key Results
- Quantitative Assessment: MFD-md consistently produced the lowest Root Mean Square Error (RMSE) for SCA calculations across all four artificial terrain types compared to the D8 (SFD) and MFD-Quinn (fixed-exponent MFD) algorithms. The RMSE for MFD-md was significantly lower, especially for the ellipsoid surface (6.09 vs. 65.76 for D8 and 6.60 for MFD-Quinn).
- Qualitative Assessment: Application to a real watershed showed that MFD-md produced a smoother and more realistic spatial distribution of flow accumulation compared to D8. While MFD-Quinn also produced smooth results, MFD-md showed a better representation of flow in valley areas, where it produced smaller flow accumulation values than MFD-Quinn, indicating a more accurate modeling of convergent flow.
- Terrain Adaptability: The results suggest that MFD-md is more suitable for convex slopes and can model various terrain conditions better than D8 and MFD-Quinn.
Limitations and Future Work
- Parameter Sensitivity: The authors acknowledge that the specific settings for $e_{min}$, $e_{max}$, $p_l$, and $p_u$ in the flow-partition function are somewhat arbitrary and were chosen to illustrate the adaptive approach. A detailed study on the impact of these parameters on the algorithm's usability and robustness is suggested for future work.
- Other Topographic Attributes: While the maximum downslope gradient proved effective, the authors suggest that other local topographic attributes, such as plan curvature and profile curvature, could be explored in conjunction with the maximum downslope gradient to further refine the adaptive flow-partition exponent model.
- DEM Pre-processing: The paper mentions the necessity of pre-processing DEMs to handle depressions and flat areas, a standard practice in flow-routing studies. The specific method used (Martz and Garbrecht's) is referenced but not detailed, implying that the performance of MFD-md could be influenced by the chosen pre-processing technique.
Summary updated