Impact of image compression on CNN performance metrics for CPS nodes at the Arctic Tundra

In cyber-physical systems (CPS), data can be col-lected and pre-processed at the edge nodes before being sent to a back-end server. However, for environments like the Arctic tundra, both energy and data networks are constrained. The nodes must save energy to have long operational lifetime from a single battery charge. Consequently, the size of the data to be sent should be reduced as much as possible without losing the information needed to perform the analysis at the back-end. This paper evaluates the effects of data size reduction at edge nodes on the analysis performance at a back-end server. The use-case considers edge nodes deployed in the Arctic tundra to take pictures of animals. These images are then transmitted to a back-end server to determine the species of the animals using a convolutional neural network (CNN). We identify 17 functions for reducing the data size of an image series. We run several combinations on a series of images collected on the tundra. Our experimental results show a possible reduction in the number of bytes required to represent the images between 90.2% and 92.1% with decrease in mean precision and mean recall by less than 0.03. Re-scaling images is required to reach large size reductions on the image series.