The overall accuracy of the model was determined to be 59.375 %, meaning that approximately 60% of all 2015 fire events that occurred in high susceptibility areas were predicted by the susceptability model (Figure 21). This indicates that the model was fairly accurate when determining high susceptibility areas, with a majority of fire events being correctly predicted. The unpredicted fire events were mostly smaller fire events, with larger fire events being predicted with a higher precision (Figure 21). These smaller fire events represent areas where fires occurred but did not spread very far, meaning that our analysis of susceptibility did indicate that higher susceptibility areas were more likely to produce larger fire events. This variation in size of fire events is definitely a strength of the analysis, and indicates that there is value to the susceptibility analysis overall for determining the likelihood that a fire becomes a major issue. Smaller fire events indicate fires that did not spread as wide as larger fire events. The fact that these were harder to predict than larger fire events shows that the model was accurate at predicting susceptibility, due to the fact that these smaller fire events would have occurred on lower susceptibility land and would not be able to grow and turn into more extensive fires.
Figure 21. Fire Events that were Predicted and Not Predicted by The Susceptibility Model
Many sources of error in the model could have attributed to this predictability rating, including the low sample size of the fire event data and the inconsistent temporal scale of the input data. There is also a large amount of randomnes that is associated with fire events in general, which is most likely the primary factor that affects the predictability of the model.
There was also only 32 pixels in the fire event dataset that were symbolized as 2015 fire events, which represented an area of 22.205 km2. This indicates that approximately 13.32 km2 of burnt land was correctly identified, and 8.882 km2 was unidentified. This low amount of pixels that were available for accuracy assessment resulted in a low sample size to test the model on. This small sample size could have resulted in an accuracy test that was inaccurate overall, and using a fire event dataset with more fire occurrences could have potentially resulted in a different error rating.
The susceptibility model also did not only use 2015 specific data, since it could not be found for every factor in the study area. 10 year climate normals were used instead, which introduced a high level of error into the final susceptibility analysis. In order to map fire susceptibility more accurately, short term fire weather forecasts would need to be used (Zhang et al., 2017). This is due to the fact that climate and weather can vary drastically over time and create unpredictable results that can potentially create unpredictable fire events (Zhang et al., 2017). This compounded with the unpredictability of human caused ignitions creates a large amount of error in the analysis (Zhang et al., 2017).
Human caused fire events do not occur in a highly predictable manner, with many unknown factors being involved (Zhang et al., 2017). This is in large part due to the fact that ignition events are hard to predict, since many of them occur from highly unpredictable events such as human caused ignitions and natural caused ignitions (Guo et al., 2016; Kilinc and Beringer, 2007; Valdez et al., 2017). The susceptibility model took into account these factors as best as could have been done using the data that was available, but there is no way to predict these ignition events with high precision (Valdez et al., 2017). As a result of these factors, the fire susceptibility model was accurate overall, but could not predict fires with extremely high accuracy.
