Discussion

Overall, the MCE and least-cost path did not work as well as expected to identify regions of hazard within backcountry terrain in the study areas. Areas of success will be highlighted along with the mentioning possible areas of improvement for future models. Within relevant academic literature (Chobin et al., 2019) the applications of machine learning have added in solving some of the challenges that were faced during this project.  The applications of neural networks and machine learning may be the next step for this project.  

Hazard Identification

The MCE that was produced successfully identified regions of less hazard. Given the input parameters, the model did a good job identifying slopes that were suitable.  While it was able to locate regions that were less hazardous than others. It appeared that there were more suitable regions on the map that it could have selected. There may be a multitude of reasons for the failure to select the most suitable region, it could have been that the total area was too large so the locate tool had to comprise when it was selecting the larger. Additionally, when comparing input factors to the outputted hazard map, the AHP could have been skewing the results too far in one direction. By using a raster that had a lower resolution, some of the sensitivity within the curvature function may have been lessened.  

Within the identification of hazards and finding a suitable zone, the size of the raster may have had an impact on the result. A raster size of 25 meters, may have been too fine of a resolution. The curvature function may have identified regions that were either concave or convex but were too small to have an impact on the avalanche hazard. Along with placing a high weight in the AHP on curvature, an overly sensitive layer may have skewed the results.  

Descent Path Identification

The identified descent paths had a tendency to stray from the identified regions of least hazard. The deviation may have been two-fold. Firstly, the least-cost path tool has the ability to select individual cells to travel to whereas the locate tool was constrained by a general shape while identifying suitable locations. Because of the least-cost path’s freedom, it may have been able to identify a safer, yet smaller sliver than was identified in the locate tool from the MCE. Secondly, the least cost path selection may have been impacted by the weights applied by the AHP that will lead to deviation from the identified area of the lowest hazard.  

Additional impacts on the least-cost path tool may have been the cell size of the raster. Like the locate tool, an overly sensitive strongly weighted curvature may have had an impact on the routes identified by the model.  

Further Research

Further work could be done to identify the importance of forest density on avalanche hazards within the study regions. There is currently a limited amount of high-resolution LIDAR data available to calculate forest density within the study region. As the government of Canada continues to update its High-Resolution Digital Elevation Model (HRDEM) with LIDAR data, it may be a possibility in the future to update the study with LIDAR data. Additionally, the addition of meteorological factors on avalanche risk in the study areas will complete the overall risk assessment of the region in regards to avalanche hazards.  

As discussed above, the implementation of machine learning in future projects may aid in a successful outcome.  A neural network, when fed the right data, may be able to identify correlations between inputs and reweight factors to achieve a more suitable model. 

Conclusion

The study of physical factors and their impact on the avalanche hazard of three different regions in British Columbia yielded mixed results. While the MCE was able to identify the regions that were of lesser risk, it appeared that there were better locations that were not selected in all study areas. Additionally, the least-cost path did indicate the least hazardous descent from the point within the identified region to the nearest road, but at times it did stray away from the greater area of least hazard.  

As the computational representation of snowpack and avalanche analysis improves, human judgment is still the best form of analyzing terrain and choosing routes. Snow science intertwined with terrain makes perfect avalanche prediction near impossible due to the complex factors involved. Therefore it is recommended to approach avalanche terrain with conservative risk assessments based on desktop studies and field observations.