Epsilon Archive for Student Projects

Abstract

Estimation of cervid forage cover is one of the tasks of forest resource management. It improves the browsing damage prognosis and allows for planning more precise mitigation strategies. In this thesis, I compared the accuracy of four different types of models in predicting the percentage cover of cervid forage. I used data on the Swedish Laser Scanning survey (SLSS), climatic variables (annual temperature and precipitation) from the wordclim database and tree species volume proportions from SLU species maps to train models. The Swedish National Forest Inventory (NFI) was a source of the data about forage cover. Canopy height, canopy cover and elevation were either taken from or calculated based on data from SLSS. I fitted two generalized linear mixed effect models with a beta distribution, one generalized linear additive mixed effect model and one random forests model with forage cover of Scots pine (Pinus sylvestris), oak (Quercus robur) and birch (Betula pendula and B. pubescens) as the response variable. Results varied both among species and methods, but Random Forest was the most accurate model for all species while GAMM performed the worst. The pine models achieved the best r2 values, but r2 values were relatively low in all cases. This suggests that in addition to the height of the canopy, canopy cover, species composition, mean annual precipitation, mean annual temperature and elevation, other predictor variables may be needed. Future studies creating predictive models for the percentage cover of these forage plants should utilize additional predictor variables.