Comparing methods of addressing uncertainty in urban forestry damage assessments following Hurricane Irma using random forest and k-nearest neighbors

Background/Question/Methods

In September 2017, Hurricane Irma made landfall in Florida, with heavy rains and wind speeds up to 146 kph. Urban forest damage was widespread and included uprooted and downed trees, limb loss and defoliation. When performing post storm inventories in urban forests it can be difficult to determine the particular source of observed damage due to the active and individualized management they undergo. To combat this, researchers often require land managers and residents to confirm the origin of any recorded damage. Unfortunately, such confirmation is not always possible, causing elevated rates of data uncertainty. Utilizing Random Forest (RF) and k-Nearest Neighbors (k-NN) algorithms we tested our pre- and post-Irma urban forest inventories from central Florida to determine how different methods of handling uncertain damage observation impacts overall model determinations. In one set of tests, we investigated methods to impute uncertain data, as compared with models where uncertain data were excluded, comparing variability in model performance. Storm damage was imputed with previously established species wind rating and storm damage likelihood determined by field crew observation. In a second set of tests, we compared the performance of models using only confirmed observations, simulating varying rates of storm damage.

Results/Conclusions

Long established relationships between windstorm damage susceptibility and factors such as tree diameter, individual crown condition and plot density were supported in both methods of imputation tested. However, there was wide variation among models in terms of factors which were determined to be important in predicting damage, such as location, plot species composition, and land use. When only confirmed data is considered, crown condition prior to the storm was consistently ranked as an important factor in predicting damage regardless of the simulated rate of damage, while the variability of tree heights and plot location were consistently ranked lowest. However, the importance of factors such as plot basal area and land use varied with the tested rates of storm damage. Given the risk to urban communities that tree failure can pose during severe climate events, it is important to determine consistent and appropriate risk factors that are robust to imperfect data. Given the variation found in our model tests, we recommend that future analyses of urban and suburban forest not only consider potential uncertainty, but also use cross validation to achieve results with robust testing methods.