Neural networks and satellite images-based shrub tundra landscape study: phenomena with fuzzy geometric and categorical boundaries

Many studies using convolutional neural networks, including in the field of satellite images, are aimed at recognizing clearly defined objects, such as cars or individual trees. Here we present the first results on mapping the growth stage of tundra shrubs, which is a ‘‘fuzzy’’ target for a network: there are no obvious geometric boundaries of the growth stages, as well as a clear division between the classes representing the growth stages. For six high-resolution satellite images of three Low Arctic landscapes, we achieved an F-score of 0.8-0.9, what is very promising for further ecological research. We also found that different landscapes and stages of shrub growth are not equally error-prone.