Discussion

Model and Variable Characteristics

The prediction-minded evaluation of data/variable types reported in the previous section states the obvious—that more independent variables tend to improve model predictability of dependent variables. This situation says nothing about the direct applicability of all the work involved with this study and the possible creation of predicted damage values across or outside of the study area, or both, with the various models produced. What is of importance in this work, however, is where, with respect to the variables and variable types used, the gains in model fit occur, although no statistical inference can be associated with these gains.

We expect that similar models can yield predicted canopy changes with RMSE values at or near 0.13 (13 percent) for situations like the passage of a strong hurricane over a mostly undisturbed Southern forested area, such as south Mississippi before Katrina. These estimates can be improved from the forest industry perspective. Industry is often focused on the softwood resource in the South, which is where the best of the stratified models we developed demonstrated a RMSE 0.09 (9 percent). Model fit was comparable in mixed and softwood stands but was poor in hardwoods. This could be the result of a variety of issues from some unknown data bias that was unintentionally introduced into the modeling process or some natural occurrence unknown and unaccounted for in these analyses. These poor results could also illustrate the inherent difficulty and complexity in modeling conditions in hardwood areas.

Potential model flexibility to create comparable predictive models, regardless of the use of poststorm imagery, was a much sought after finding in this study with mixed results. The reason for this exploration was to display the applicability of modeling anticipated storm damage prior to a weather event. This focus was best explored in comparing the other models (or adjusted) pooled fit values for the MIFI and prestorm variables models and Hocking method pooled MIFI, prestorm, and poststorm variables models where RMSE values were 0.153 versus 0.130, respectively. This comparison is somewhat indecisive as MIFI, prestorm, poststorm variables models outperformed the MIFI and prestorm variables models but only by a small amount (difference in RMSE of < 0.03). Similarly, in the corresponding 8 variable models, there was a difference of 0.02 with respect to RMSE. This difference is of a smaller magnitude, however, than the lack of MIFI data comparisons of overall pre- and post-storm variables (R²_adj=0.492 and RMSE=0.172) versus prestorm only variables (R²_adj=0.277 and RMSE=0.205). In comparing the Hocking pre- and poststorm model (RMSE=0.172) versus the adjusted pooled MIFI and prestorm model (RMSE=0.153) and the corresponding 8 variable models (RMSE=0.181 versus RMSE=0.190), it does appear that use of the MIFI data in model development at least offsets, maybe even improves, model performance in using prestorm data with the absence of poststorm data.

Possible Model/Variable Improvements

Actual field damage values are being collected in MIFI’s Southeast region, which includes: Jefferson Davis, Covington, Jones, Wayne, Marion, Lamar, Forrest, Perry, Greene, Pearl River, Stone, George, Hancock, Harrison, and Jackson counties. These data are the direct metric of interest in this series of work, as opposed to the photo-interpreted canopy metric used here. Incorporation of these data is expected to improve development, although model fits may worsen, of any hurricane damage assessment model subsequently created due to the dependent variable’s added meaning. The data could also help address a noted problem of hardwood defoliation versus damage. Poststorm high resolution imagery indicated that many hardwood areas, particularly in the Pearl River bottom, were defoliated with only minor damage to tree crowns and boles. Differences in this defoliation versus damage aspect of hardwood areas may also be more sensitive to individual hardwood species, which is one of the field metrics, as opposed to the whole hardwood type.

Along with the analysis of field data, future work will also incorporate statistically inferential results, such as variable significance, as opposed to the simple fit comparisons made in this work. These analyses will provide more meaningful results with potential adaptations for collinearity and validation of model assumptions. Model development for repeated application may also be achieved in order to create a more robust and possibly automated product.