Detecting Reindeer Carrion in the Arctic Tundra of Svalbard using a Convolutional Neural Network

By Frank Fluit

Abstract
There is an increasing call for quantifying carrion biomass production. Having more information available about the production of carrion in an ecosystem leads to a better understanding of foodwebs and overall ecosystem functioning. Combining information of single carrion with the amount of carrion in a region the biomass production can be assessed. With the use of UAV and CNN technology this study determined the possibilities of detecting carrion in the Adventalen region in Svalbard Norway.

This was investigated using the MobileNet v2 CNN architecture. Transfer learning was applied based on the Imagedata set in two different settings. Only the training of the top dense layer was investigated and a finetuning approach was taken where the top 100 layers of the model were set to trainable. Various training scenarios for detection and classification of reindeer carrion were designed. Other aims of the research were to look into the effects of data augmentation techniques, the use of various carrion: no carrion ratios in the training data and the effect of using centered and non-centered training data. The results were evaluated using the accuracy, precision and the recall metrics. Then was reflected on how these would result to a real life application.

The research indicated an effect of data augmentation, but no direct positive effect on the results was found. It was also found that training on uncentered images typically obtained better results on an unbalanced uncentered test sets, obtaining recall values above 0.8 and acceptable precision values. These results were deemed satisfactory for a real-life application, however exploring heavier architectures and cross-validation could improve the generalizability and the metric scores. The classification results however were not deemed high enough for a real-life application.