Date of Award

Summer 8-22-2020

Level of Access

Open-Access Thesis

Degree Name

Master of Science (MS)

Department

Forest Resources

Advisor

Parinaz Rahimzadeh-Bajgiran

Second Committee Member

Aaron Weiskittel

Third Committee Member

David MacLean

Additional Committee Members

Brian Roth

Abstract

Insects are one of the most significant agents causing landscape level disturbances in North American forests, and among them, spruce budworm (Choristoneura fumiferana; SBW) is the most destructive forest pest of northeastern Canada and U.S. The SBW occurrence, its damage extent and severity are highly dependent on characteristics of the forests and availability of the host species (spruce (Picea spp.) and balsam fir (Abies balsamea (L.) Mill.)). This study developed novel methodologies to detect and classify SBW defoliation and to map SBW host species using remote sensing techniques. Optical multispectral remote sensing satellite imagery presents a valuable data source for regional-scale mapping of forest composition as well as defoliation severity and can be effectively used for monitoring insect outbreaks. This study developed two separate models to map both the distribution and abundance of SBW host species as well as the severity of defoliation at 20 m spatial resolution utilizing Sentinel imagery. The two models were integrated to effectively monitor the SBW defoliation. For the detection and severity classification of SBW defoliation, we used Sentinel-2 imagery and site variables (elevation, aspect, and slope) and compared the capabilities of various spectral vegetation indices (SVIs), in particular red-edge SVIs, to detect and classify SBW defoliation using Support Vector Machine (SVM) and Random Forest (RF) models. The study was carried out in the Northern part of New Brunswick, Canada. Results showed the superiority of RF in model building for defoliation detection and classification into three classes (non-defoliated, light and moderate) with overall errors of 17% and 32%, respectively. The most important Sentinel-2 based variables for the best model were Inverted Red Edge Chlorophyll Index (IRECI), Enhanced Vegetation Index 7 (EVI7), Normalized Difference Infrared Index 11 (NDII11), Modified Chlorophyll Absorption in Reflectance Index (MCARI), and Modified Simple Ratio (MSR). Elevation was the only site variable significant in the final model. The study concluded that red-edge SVIs were more effective variables for light defoliation detection compared to the traditional SVIs such as Normalized Difference Vegetation Index (NDVI) and EVI8. These findings can help improve the current remote-sensing based SBW defoliation detection techniques. For SBW host species classification, Sentinel-1 synthetic aperture radar (SAR) and multi-spectral Sentinel-2 imagery were used in combination with several site variables (elevation, slope, aspect, topographic wetness index, soil types, projected climate site index for year 2030, and improved Biomass Growth Index (iBGI)). The study was carried out in the same location where the first study was conducted but extended to a larger area (northern parts of New Brunswick, Canada) using a total of 191 variables. We found Sentinel-2 time series in combination with single spectral bands and spectral vegetation indices (SVIs) promising to map SBW host species using a RF algorithm, with an overall accuracy (OA) of 71.34% and kappa coefficient (K) of 0.64. The use of Sentinel-1 SAR data alone with elevation showed a decent result (OA: 57.5 and K: 0.47). Furthermore, the combination of Sentinel-1, Sentinel-2 and elevation provided us with an OA of 72.3% and K of 0.65. The most important Sentinel-2 variables for the best model were from the images of late spring and fall seasons including three single spectral bands and seven SVIs mostly from near-infrared, red-edge and shortwave-infrared regions. Prediction of spatially explicit SBW host species data is essential for identifying vulnerable forests, tracking the SBW defoliation and minimizing the forest loss as well as serving as a vital input for modelling and managing insect impacts at the landscape and regional scales.

Share