Q15 – S. Donovan

Using digital hemispherical imagery and EO-1 Hyperion hyperspectral data for spruce budworm defoliation forest health surveys

A significant forest health issue in Canadian forests is the damage caused by periodic spruce budworm (Choristoneura fumiferana Clem.) outbreaks, which covered 52 million hectares at the peak of the last outbreak in 1975. The current outbreak has already reached 7.2 million hectares as of 2017 aerial surveys in the province of Quebec and populations are low but increasing in neighbouring New Brunswick. Quantifying the annual extent and severity of defoliation is an important input for forest management tools such as the spruce budworm decision support system (SBW DSS). Estimating defoliation over large forested areas currently relies on aerial surveys using broad defoliation classes such as light, moderate, and severe. This method is subjective, inconsistent across observers, and prone to errors temporally and spatially. Plot-level branch and ocular sampling methods are more accurate than aerial surveys but only feasible for covering small areas of interest. Objective one of this research investigated the use of repeated digital hemispherical imagery within plots quantifying the gap fraction changes to estimate annual defoliation caused by spruce budworm. The second objective examines temporal EO-1 Hyperion data for classifying annual defoliation using a network of 75 plots with accurate branch defoliation data for validating indices changes.

For the first objective, hemispherical images were collected for 75 plots in Gaspé, Quebec before and after defoliation periods in 2015 and 2016 with accompanying annual branch and ocular plot defoliation estimates. Nine average plot gap fraction change metrics derived from the hemispherical images for each plot in each year, along with an insecticide spraying and 20 plot attribute variables comprised the dataset. Gradient Boosting Machine analyses and pairwise collinearity diagnostics identified insecticide spraying, gap fraction change May-October, and balsam fir (Abies balsamea L. Mill.) % basal area as the most important explanatory variables for annual plot defoliation. Using Logistic Generalized Linear Models and Random Forests stepwise modelling, results indicated defoliation predictions on one-third of the plots for model testing had a root mean squared error (RMSE) of 14-22% when including the explanatory variables spraying, gap fraction change May-October and balsam fir % basal area. When excluding the spraying variable, the RMSE was higher at 18-24%. We concluded that quantifying gap fraction change from hemispherical photos is a feasible, non-destructive, and objective ground-based method to assess canopy foliage changes caused by spruce budworm.

Figure 1. Relative influence (variable importance) for the top 10 out of 30 explanatory variables using the combined year dataset (n = 140) to predict annual branch defoliation per plot (%) for a) including, and b) excluding the Spray variable. Variables with the same superscript number were multicollinear at a correlation coefficient threshold of r ≥ 0.7.

 

Figure 2. Annual defoliation predicted using four models plotted against measured annual branch defoliation for 44 retained validation plots, for 2015, 2016, and combined data: a) GLM and b) Random Forests (RF) models using Spray, GFC_MO, and BF_BA% explanatory variables; c) GLM model using GFC_MO and BF_BA%; and d) RF model using GFC_MO, BF_BA%, and GF_October. LOWESS regression lines are fit per year and for combined data, and solid coloured symbols represent sample plots affected by insecticide spraying.

 

The EO-1 Hyperion satellite has shown promising results with a 15% accuracy increase in mapping spruce budworm defoliation validated using coarse-scale aerial defoliation survey data versus using multispectral data (Huang, 2015, MScE thesis at UNB). Our second objective is investigating vegetation indices change detection derived from EO-1 Hyperion hyperspectral data and validated using accurate plot-level defoliation data from the 75 plots in 2016. Maximum likelihood and Support Vector Machines supervised classifications are being examined for pixel-based mapping of annual defoliation in Hyperion scenes. Additionally, a pansharpening technique to increase spatial resolution using methods developed by Dr. Zhang’s research group at UNB is being investigated to compare mapping accuracy improvements.

Currently, progress has completed preprocessing of the EO-1 Hyperion data and is examining vegetation indices changes for pixels overlaying defoliation plots in order to determine the sensitivity of each indices in relation to annual defoliation. Future work will involve classifying the remaining scene pixels and producing defoliation severity maps.

 

Conference Presentations

Donovan, S. and D.A. MacLean. 2017. Quantification of spruce budworm defoliation using hemispherical imagery and EO-1 Hyperion hyperspectral satellite data. Pp. 30–37 In: Proc. SERG International 2017 Workshop. Feb 6–9, 2017, Fredericton, NB.

Other Presentations

Donovan, S. and D.A. MacLean. 2018. Quantification of forest canopy changes caused by spruce budworm defoliation using digital hemispherical imagery. Northeastern Forest Pest Council 80th AGM, Burlington, VT, Mar 13–15.

Donovan, S, and D.A. MacLean. 2018. Using EO-1 Hyperion satellite and digital hemispherical imagery for spruce budworm forest health surveys. Enhanced Inventory Project – National Overview and NB Team Meeting. Fredericton, NB, Mar 20–21.

Donovan, S. and D.A. MacLean. 2017. Quantification of spruce budworm defoliation using hemispherical imagery and Hyperion hyperspectral satellite data. NSERC AWARE Network 2nd AGM, Edmundston, NB, May 16–18.

Donovan, S. and D.A. MacLean. 2017. Spruce budworm/forest health assessment. Enhanced Inventory Project – National Overview and NB Team Meeting. Fredericton, NB, Jan. 10–11.

Donovan, S. and D.A. MacLean. 2016. Quantification of spruce budworm defoliation using hemispherical imagery and Hyperion hyperspectral satellite data. NSERC AWARE Network 1st AGM, Corner Brook, NL, May 24–26.