Utilizing transfer learning with artificial intelligence for scaling-up lichen coverage maps

Lichen mapping is essential for sustainable caribou and lichen conservation. Previous studies have used artificial intelligence to create lichen coverage (%) maps using a scaling-up methodology. This paper proposes a transfer learning approach to lichen mapping, where the model weights from a lichen coverage (%) neural network trained on Quebec and Labrador data were used for training a northern Ontario model. The training and evaluation dataset consisted of 22 10-m lichen maps from Geraldton, Martin Falls, and Peawanuck Ontario, aligned to Sentinel-2 imagery. The model with transfer learning outperformed a similar neural network with randomized initial weights and a random forest model, despite predicting lichen in different ground conditions than Quebec and Labrador. A northern Ontario lichen map was created using a Sentinel-2 mosaic, displaying high amounts of lichen surrounding Peawanuck, Ontario. Whenever feasible, transfer learning approaches should be considered when organizing regional vegetation coverage (%) maps.