Remote Sensing of Boreal Coarse Woody Debris: The Wrap-Up

My MSc thesis, entitled Remote Sensing Boreal Coarse Woody Debris, was successfully defended on September 13th, 2019. The main objective of the research I conducted was to assess the accuracy and feasibility of using computer vision techniques and modelling to detect and measure dead trees (i.e. coarse woody debris, CWD) in piloted-aircraft data of boreal forests. The two following peer-reviewed publications, also released as chapters in my thesis, spawned from this research.

But wait! The events I just described mark the end of my journey as an MSc student… Read the following sections of this post for an account of my main learnings, interesting discoveries and experiences I had along the way.

Mapping Coarse Woody Debris with Random Forest Classification of Centimetric Aerial Imagery

Brief description

We developed novel methods for mapping boreal CWD on piloted-aircraft imagery using machine learning and object-based analysis. The presented methods achieved from 80 to over 90% accuracy, depending on whether training samples were in the application area.

Graphical abstract

Citation

Queiroz, G. L., McDermid, G. J., Castilla, G., Linke, J., & Rahman, M. M. (2019). Mapping Coarse Woody Debris with Random Forest Classification of Centimetric Aerial Imagery. Forests, 10(6), 471.

Estimating Coarse Woody Debris Volume Using Image Analysis and Multispectral LiDAR

Brief description

Novel models for estimating CWD volume as well as the first ever extensive (4300 hectares) high-resolution maps of CWD volume are presented in this paper. Our models achieved good accuracies (0.623 R², 0.224 RMSE m³/100m²) even where CWD was invisible in the input images due to superimposed vegetation and shadows.

Graphical abstract

Citation

Queiroz, G. L., McDermid, G. J., Linke, J., Hopkinson, C., & Kariyeva, J. (2020). Estimating Coarse Woody Debris Volume Using Image Analysis and Multispectral LiDAR. Forests, 11(2), 141.

Caribou Simulator

My master’s thesis was originally entitled Remote Sensing of Coarse Woody Debris in Caribou Habitat. Since the focus of my research was on remote sensing methods and not caribou, the title ended up changing to what it is now. Nevertheless, I read and researched some literature on caribou habitat in the context of Alberta's boreal forest, which is home to some of the largest oil reserves in the world. I was fascinated with the history and narratives on the decline of caribou herds in the province. It is hard not to have sympathy for these animals. If the reader is interested in this topic there is a great documentary by CBC on caribou.

I occasionally participate in game-making marathons (i.e. game jams), where a game is completely developed from scratch in 48h using a theme that is unveiled at the start of the event. I participated in such an event at the end of my first semester in grad school, December 2017, when I was reading a lot of literature on caribou. Thus, Caribou Simulator was created, a cute puzzle game where you lead a caribou herd through a dangerous part of the boreal forest.

Click on the image to see a video clip of the Caribou Simulator game!

Field Season

At the end of June, 2018, I went to a study area in northeastern Alberta, close to the small settlement of Conklin, to collect CWD measurements in the boreal forest for my Master’s research. The field season comprised of a total of 22 field days, with a 10-day break at the beginning of July. I had a field assistant, Alexandria Nicole Hamilton, to help me collect my CWD data; as well as two other student companions, Sean Sugden and Keifer Biddle, who were collecting data for a different research project.

The study area has a rolling terrain, with dry uplands mainly composed of coniferous, deciduous and mixed forest stands; and wet lowlands mostly composed of black spruce fens as well as swamps, bogs and marshes which are mostly shrubby and mossy. I visited a total of 108 sites and surveyed 945 CWD pieces.

One of the most memorable aspects of the forest, which me and all my field companions noticed and loved talking about, was how soft and puffy the mossy ground could be. It was so soft you could literally have a comfortable sleep on it like it was a mattress. We encountered it all over the lowlands and sometimes is was so tall you would sink into it and feel like you were walking over a mountain of cotton candy.

Last but not least, we had a scary close encounter with a bear while in the forest. The day this photo was taken we were distracted collecting data on difficult terrain; meanwhile a big black bear was joyfully sniffing and following the trail of a delicious pineapple treat I forgot was in my backpack. Us and the bear had an interesting exchange and then we each went our way. No more data was collected that day.

GEOBIA

The approach I used to detect CWD on aerial images in my first paper is known as Geographical Object-Based Image Analysis (GEOBIA). It consists of segmenting the aerial images into polygons which contain relatively homogeneous pixels and then studying these polygons according to their spectral (pixel values), spatial and relational characteristics. It is a sophisticated approach, especially useful for dealing with high-resolution imagery. It is unfortunate that currently there are limited options for applying this technique aside from the eCognition application from Trimble. I predict that soon there will be many viable open source alternatives for GEOBIA. Orfeo toolbox and RSGISLib are examples of existing open source solutions that have limited GEOBIA capabilities.

Machine Learning

The most common type of classification in GEOBIA involves the use of human-created rules to separate the image-objects (polygons) into different groups. However, I found it difficult to obtain good accuracies using this approach to map CWD, considering that my object of study has high variability depending on the tree species, decomposition and surroundings. I turned to machine learning (random forests), an approach where a training sample is fed to an artificial intelligence which then creates a very sophisticated set of classification rules. I found the pros of machine learning to outweigh the cons in this context. Creating a large training sample was as simple if not easier than creating human-based rules, and the outputs were much more accurate.

Multispectral LiDAR

In my second paper I selected models for estimating CWD volume, and I found that Multispectral Light Detection and Ranging (ML) has good potential to provide information near the ground even in areas with closed canopy. This was exciting for a couple of reasons, firstly it had the potential to improve my models, and secondly ML is an emerging technology that has not yet been studied in this context (CWD) so it presented a gap in the literature I could fill. My results suggested that ML has great potential, yet it needs to be further studied in terms of calibration to actually improve CWD models.

Final Thoughts

Throughout my journey from September 2017 to September 2019 I experienced much and made many new friends. I even got engaged and became a father! But those are stories for some other time… I feel very fortunate to have been able to see, learn, accomplish and share all these things during this time. I am immensely grateful to all the people who have provided the support I needed: family, friends, colleagues and my supervisor. I know that everyone’s experiences are different, and that graduate school is not for everyone, but if you ask me, I would recommend taking this sort of challenge as there is more to education than just knowledge…