Tree ID in Skattebo

Glade, BC

Intro

In the Southern Interior of British Columbia, a narrow trail weaves through the Skattebo forest, contouring the Kootenay river; it connects a small ferry-only-access community of Glade to Castlegar. The Skattebo Educational Forest stretches to the south of Glade Creek. It is a valuable community resource that provides access to nature trails, numerous beaches, and hunting and gathering opportunities. The land was transferred from the Cominco to Selkirk College in 1999 and has been used for research and other educational purposes since.

Using drones to map a portion of the Skattebo Educational Forest

Credit: B. Wilson

Hi, my name is Vadim and I’m one of the founders of Above Sensing Ltd. and a forever student. Because of my education at Selkirk College, I fly drones and use the data they provide to improve the ways we manage our resources. Drones or unmanned aerial vehicles (UAV) or remotely piloted aerial systems (RPAS), a jargon often used by Transport Canada, are disruptive technology that is continuously introduced to many aspects of our lives. Drones perform many useful functions and are not simply limited to photography and videography, which are commonly known uses. Drones can carry multiple types of sensors for various applications in natural resources, water, and land management.

Using UAVs to augment forestry work is an emerging trend. Drones provide access to otherwise hard-to-reach areas making work safer and faster. Tree species identification, their attributes, and forest spatial structure are valuable characteristics for forest inventory, stand health assessment, harvest planning, ecosystem analysis, and wildfire behavior modeling. A change of perspective from the above allows for decision-making based on geospatial analysis without leaving the office, as well as providing the additional benefit of being better prepared for fieldwork.

You may have privacy concerns about drone operations. Drones are classified as aircraft; given the pilot’s certification level, they can utilize most of the airspace just like any regular airplane or helicopter. That means it is not illegal to fly over your property. However, drone operators must follow the general privacy guidelines outlined in the federally regulated Personal Information Protection and Electronic Documents Act (PIPEDA). It’s worth noting that British Columbia has its own stricter privacy laws which are based on the PIPEDA. However, privacy laws do not apply when the purpose of an aerial survey falls under journalistic, artistic, literary, or academic purposes. Nevertheless, it is essential to be open about ongoing projects and share information with the community if potential concerns arise.

The Research

The primary goal of my research is tree species identification from the data collected by a drone with a multispectral sensor. Multispectral sensors record how much light gets reflected in specific light spectrum bands. Instead of a “normal” composite RGB picture, six monochrome images are recorded. Then, the data is analyzed and stitched together using the principles of photogrammetry by a software suite.

Processing large amounts of UAV-derived data is challenging for the human eye. Machine learning (ML) or deep learning (DL), a subset of artificial intelligence (AI), thrives on large datasets and finds patterns in complex data very efficiently. Deep learning mimics brain neuron connections by connecting data points and learning data patterns. The concept has existed since the 1950s, but it did not receive wide adoption until recently when high computing power became abundantly available.

My project explores a method of tree stand classification by developing a machine-learning model capable of identifying the common tree species in the Skattebo forest and ecosystems alike. The Skattebo forest is very diverse. The tree species include Douglas-fir (Pseudotsuga menziesii), western redcedar (Thuja plicata), western larch (Larix occidentalis), paper birch (Betula papyrifera), grand fir (Abies grandis), western white pine (Pinus monticola), and lodgepole pine (Pinus contorta) – The Kootenay Mix.

WingtraOne Taking-off

Credit: B. Wilson

I have captured two datasets at Skattebo using a fixed-wing vertical take-off and landing drone, WingtraOne GEN II, the one we use for the business. The difference between fixed-wing and multirotor drones is how they perform in the air. The fixed-wing aircraft can maintain a flight longer because its wings create lift. But unlike the multirotor drones, they cannot turn on a dime and require a wide turn-around maneuver to get back to their mission’s grid.

The drone is equipped with a post-processing kinematic (PPK) module that records additional positioning information. After the drone has landed, the accuracy of data can be improved by combing the PPK and base station data. The resulting accuracy of the Skattebo dataset is an impressive +/- 3cm in the horizontal plane. That means that the objects in the imagery are within 3cm from their actual location on Earth.

Ground control point and the accuracy of the collected data.

Collecting Individual Tree Locations

Throughout the winter I was developing the tree species identification model based on data collected during the fall. To verify my findings, the next step required collecting the ground truthing data that will be used to verify how well my algorithm predicts the tree species. To do that, I needed to manually record a few hundred points with a handheld GPS device to compare them against machine learning predictions. After a few nature walks with friends, I collected enough data to proceed with the experiment.

The problem with machine learning is the perpetual need for large datasets to be trained on. In real-life applications collecting such large samples would be inefficient and we might as well continue forest-surveying the old way. To overcome the challenge, I used data-augmentation techniques and produced synthetic data that effectively quadrupled the original dataset. The model welcomed the addition of augmented volumes and the outcome of the experiment produced unexpectedly high accuracy values. “Eureka, I solved remote sensing!” – I laughed and ran the model on the entire dataset.

Individual Tree Species Identification

I was always fascinated with drones and truly believe that adopting the drone/AI combo is the way of the future for many industries. As technology evolves, the limits of the equipment are pushed further and further allowing for better data and therefore more accurate representation of our world. This project demonstrates an effective application of a fixed-wing aerial platform and a multispectral payload to stratify the forest canopy and identify individual tree species using modern methods in a time-efficient manner. Though some challenges remain, for example, sensors carried by UAVs cannot see through the leaves all the way to the ground. The answer to that will eventually be found, I bet the solution will rely on predictive models and machine learning in one way or another.

This project became possible because of a collaboration between Selkirk Innovates, Mitacs Accelerate Entrepreneurship Program, ETSI-BC, and Above Sensing Ltd. Thank you goes to all my bushwhacking volunteers: Marcus Larch, Inza Maki, North Ross, and Santiago Botero. I couldn’t do it without you!