Transferability of a Mask R–CNN model for the delineation and classification of two species of regenerating tree crowns to untrained sites

Following harvest, monitoring reforestation success is a crucial component of sustainable management. In Alberta, Canada, like other jurisdictions, the efficiency of the current plot-based forest regeneration monitoring regime is challenged by the cost of accessibility and the declining availability of qualified field crews. Fine spatial resolution imagery and deep learning have been proposed as alternative monitoring tools and have proven successful under experimental conditions, yet how successfully models can be applied and transferred between a range of untrained sites and conditions remains unclear.

In this research, we repurposed a mask region-based convolutional neural network (Mask R–CNN) model that was previously trained to delineate coniferous tree crowns to instead segment instances of two species of regenerating conifers. We transferred learned parameters by replacing original single-class labels with photo-interpreted species information and retraining a selection of the network's parameters. We assessed the transferability of the new model by testing on five untrained sites, representing a range of forest types and densities typical of reforestation in the region. Results yielded a mean average precision (mAP) of 72% and average class F1 scores of 69% and 78% for lodgepole pine (Pinus contorta) and white spruce (Picea glauca), respectively, demonstrating successful transferability. We then investigated an additional transfer learning scenario by iteratively adding data from four of the five sites to the training set while reserving data from the remaining site for testing. On average, this improved mAP by 5%, lodgepole pine F1 by 7%, and white spruce F1 by 3%, and demonstrated that trained models can be continuously improved as sufficiently representative data becomes available.