Quantifying Drought-induced Changes in Green Vegetation Fraction and Classification Accuracy Using Hyperspectral Data for the Central Sierra Nevada, California

Abstract

Persistent drought conditions and associated vegetation mortality in the central Sierra Nevada of California were analyzed from 2013-2015 using a combination of field-derived polygons and AVIRIS hyperspectral data. Linear Discriminant Analysis (LDA) was used to classify hyperspectral data into five land cover classes based on dominant flora. LDA accuracies were compared across years in order to determine whether classification accuracy was correlated with increasing drought severity. It was determined that 2013 had the greatest accuracy and 2015 had the lowest. However, this trend was influenced by Bidirectional Reflectance Distribution Function (BRDF) effects in the densely forested landscape. Fractional cover data of green vegetation (GV), non-photosynthetic vegetation (NPV), and soil were obtained from the US Forest Service to analyze which land cover classes and which elevation intervals experienced the greatest fractional cover change, which are both indicators of vegetation senescence and mortality. GV loss deemed the most appropriate indicator of vegetation senescence and mortality as NPV and soil appeared to be confused by the Multiple Endmember Spectral Mixture Analysis (MESMA) method used to obtain the fractional cover images. Mixed oak woodland (MO) and mixed low conifer (LC) forests experienced the greatest and second-greatest decreases in GV, respectively. Lower elevation areas (695-1369 m) generally experienced greater GV loss than higher elevation areas (2167-2779), which coincided with both MO and LC forest classes. The MO forest class, which occurs most in lower elevation areas, was comprised of dominantly drought resistant flora and experienced the greatest GV loss during the study period (16%). Conversely, the HC forest, which occurs dominantly in higher elevation areas, was comprised of dominantly non-drought-tolerant flora but experienced less GV loss (5%). This suggests that the differences in elevation and location of vegetation within the landscape played a larger role than the dominant floras' degrees of drought tolerance. Variations in seasonal senescence may have influenced the measured loss of GV for the MO and LC classes, which contained deciduous vegetation. However, overall GV loss in all classes, even those without trees, indicates that the landscape likely experienced vegetation mortality, especially at low elevations in the MO and LC classes.