This paper focuses on the evaluation of deep convolutional neural networks for the analysis of images of natural scenes subjected to large seasonal variation as well as significant changes of lighting conditions. The context is the development of tools for long-term natural environment monitoring with an autonomous mobile robot. We report various experiments conducted on a large dataset consisting of a weekly survey of the shore of a small lake over two years using an autonomous surface vessel. This dataset is used first in a place recognition task framed as a classification problem, then in a pose regression task and finally the internal features learned by the network are evaluated for their representation power. All our results are based on the Caffe library and default network structures where possible.