How It Works
To date, the work to identify logging scars has been labor intensive and time consuming. For our Capstone project, we have built a tool comprised of set of models that identify potential logging scars, road infrastructure and deforestation in the Boreal Forest. This tool is intended as a starting point for further study of logging scars over time.
The team started by working with readily available, high resolution ( 10 meter ) images from Sentinel-2 satellites. Next, they took the results of the 2019 Wildlands League logging scars study and recreated the logging scar masks for 27 curated sites from that study using a tool called geojson.io.
To supplement this small data set, the team also used data showing small roads in the Boreal Forest, which was available from OpenStreetMap. In addition, we applied augmentation techniques to the 27 logging scars data points.
The team started by training a U-Net convolutional network with the OpenStreetMap road data, then fine-tuned the model using the augmented logging scars data set. The team used a custom loss function with equally weighted Binary Cross-entropy, F-Loss and Dice Loss. The resulting model was evaluated using IoU (Intersection over Union ) score and Dice Coefficent.
In addition, the team developed a forest detection model to highlight areas with dense forest and areas without it. This model again used a U-Net convolutional network, trained with deforestation data from the Amazon and Atlantic forests of Brazil. Again, the team relied on augmentation techniques, in this case in an effort to make the model more generalizable to other forests, especially the Boreal Forest.