Errors in watershed models

Once DEM modelling was performed (Figure 5.25), the same procedure was followed as suggested by TUFTS University (2012) for delineating watershed models. The error model was run 100 times to create 100 delineations. The polygon shapefiles were then converted to line files and assigned a value of one. These were then converted to raster datasets and combined using ArcMap. The final output from the model is a raster where the cell values represent the number of times the watershed boundary has been modelled. Thus, at cells where the boundaries occur 100 out of a potential 100 times, the probability of the watershed boundary occurring here is 100%. The model result provides an indication of the uncertainty in boundary delineation. Test A was used for this model, as it showed both high uncertainty at low and high elevations. The analysis compared the occurrence of low probability watershed delineation on the map with areas where Uview highlighted the most uncertainty between the Test A and Ref A. Test A was also modified to match Ref A in all areas of falling in high uncertainty, as identified in the Uview z-score product (areas in z-score of 1.01-). This edited version of the Test A DEM will from now on be referred to as Test A-Cor and described as partially corrected. Probability vs. Uview visualization will be discussed as occurs in Figure 5.26.

Figure 5.26 Probability and partially corrected Test A DEM

The way the probability map is setup, is such that areas with a higher probability are in a darker shade of blue and the areas with lower probability are in a lighter shade of blue.

As can be seen in Figure 5.26 the probability map areas of higher uncertainty do not always correspond with areas having lower probability. There is one large watershed / basin delineated that with E, A, B and C in with high Uview z-score values of 2.01+ clustered around E does not affect the outcomes of the basin probability. At area F, another cluster of uncertainty occurs that also has no directly observable impact on the probability of the basin. These clusters however confirm the findings of Zandbergen (2011) and Weng (2012) that uncertainty happens in clusters. At location D however, an area with some uncertainty albeit only in the 0.51-1.00 z-score category, there is a lower probability at these areas of uncertainty, however this indicates further that at key areas even a small deviation can lead to a different derived product.

When looking at the second image in Figure 5.26, three derived basins can be seen. Basin- RA, delineated from Ref A and partly hidden behind the other basins, will be treated as the reference data for use as ground truth. Basin-TA is the basin as modelled from Test A, and

A A _{B B} C D D C E F B

the partially corrected Test A basin is named Basin A-Cor. When one compares Basin-TA and Basin-RA, they do not diverge from each other at areas identified as high uncertainty by the high z-score in Test A, but instead they diverge downstream or at areas of relative low z- score. When comparing Basin A-Cor with both Basin-TA and Basin-RA, one finds a rather different product, especially at A in the various base DEMs map in Figure 5.26. Basin A-Cor closely resembles Basin-RA at the diversion from Basin-TA at B, Basin A-Cor however deviates from both the other basins at area C, creating an extra basin. The split between Basin A-Cor and Basin-TA at D is interesting, because if one compares it with the probability map at D, it follows one of the lower probability basins. When one compares all the three basins, one comes to the conclusion, that the accuracy of a DEM has a large effect on the output. This relates to the work of Zhao et al. (2009) who found that elevation and corresponding difference from a reference was the only thing that affected DEM derived products more than resolution. As can be seen with the difference between the Basin-TA and Basin-RA even when one employs a Monte Carlo simulation to derive a probability test and introduce random errors to the magnitude of its measured error, Basin-TA may still not create a delineation similar to that of the Basin-RA that was treated as more accurate. Partially correcting Test A (Test A-Cor) also produced a different basin. This again confirms the argument that spatial data users have to understand the quality of the data they are working with, as it has a direct influence on the products generated.

The overall conclusion reached is that due to the nature of basin models and DEM derived products, the quality of the input DEM will have a direct effect on the quality of the product. Therefore, whenever a DEM is used as an input device, care has to be taken not only of the resolution of the DEM, but also the quality of the DEM (Weng 2012; Zandbergen 2011; Zhao et al. 2009). Watershed/basin models appear to be sensitive to data quality in the broader array, and even minor to small deviations may produce different output products as seen in Figure 5.26. Even when the large discrepancies were corrected, the resultant basin did not follow the course of the more accurately rated LiDAR dataset. Simulation, although useful, may also still not provide the full picture, but it is one method of understanding the potential weakness of the input data used.

Uview thus provides a tool where users and producers can explore the quality of their product, both statistically through the accuracy assessment statistics that are provided, but also visually through the visualizations provided. It is also imperative that those using

datasets be cognisant of how sensitive the processing they aim to do with the data is to data quality.