First of all, I’d like to welcome our newest researcher, Doug Bolton, who started in April and will be working on Q1 in Theme 1. With Doug joining AWARE, the research is reaching a crescendo, with eighteen questions under study now. We’re also planning to collect more data in Newfoundland, New Brunswick and at two sites in Ontario this summer. It is going to be busy
Jean-Romain Roussel is doing a PhD with Dr. Alexis Achim at the University of Laval. He is working on question 8 of Theme 2 within the AWARE research project.
Jean-Romain started his education with a degree in mathematics, physics, chemistry and biology in France. Following the classical French education pathway, he then spent three years studying engineering to get his masters degree. During this time, he worked as an intern for the American Division of Forestry in Alaska for two months. After completing his internship in Alaska, he worked as an intern for the INRA laboratory in France, where he developed an algorithm for automatic knot detection in X-ray tomography images. After completing this internship, he spent a year in French Guyana where he obtained a second masters degree in ecology of tropical rain forests and worked for the CNRS laboratory on the biology of tension wood. His work focused on the special cells which enable some particular tropical tree species to recover the verticality after a disturbance.
In 2015, he moved to Quebec to work on LiDAR as a PhD student in Laval University. His current work is focused on the effect of device parametrization on the point cloud structure. His approach is hypothesis driven and consists of building mathematical models upstream to the data and then validating them with LiDAR datasets.
Jean-Romain is mainly interested in mathematics, physics and computer sciences. As a strong supporter of open-source and open-document philosophy, he has developed his own LiDAR software both for fun and for developing his PhD algorithms without the need to use proprietary software. This open source R package is freely available and can be downloaded from CRAN.
A multi-indicator framework for mapping the potential impacts of forest management activities on aquatic ecosystem services
Aurélie Schmidt is a Master’s student working under the supervision of Richard Fournier (Université de Sherbrooke) and Joan Luther (Natural Resources Canada). Her research project (Question 17a of Theme 2) aims to develop a multi-indicator framework for mapping and assessing the potential impacts of planned forest activities on selected Aquatic Ecosystem Services (AESs) provided by boreal forests of Newfoundland, Canada. To capture a range of AESs and following the Millenium Ecosystem Assessment (2005) classification system, three AESs were selected in consultation with stakeholders: clean water provision, flood control, and recreational fishing. Aurélie is investigating spatial indicators that represent the supply of AESs and assessing the potential impacts of planned forest activities on AESs. The purpose is to improve understanding of the potential impacts of forest activities on AESs at a watershed level in order to help forest managers plan and manage forest resources and ensure the sustainable supply of AESs.
A key challenge of Aurélie’s research was to define an appropriate spatial mapping unit that takes in account forest manager needs, as well as the scale of available data. For the watershed delineation, the ArcSWAT tool was used. ArcSWAT assesses the recommended hierarchical levels from a specified threshold level known as a Critical Source Area (CSA) that corresponds to the minimum upstream drainage area that is required to initiate a stream. Using ArcSWAT, the watershed was divided into 30 sub-watersheds with a mean size of 1961 ha (Figure1).
Another challenge was to select relevant spatial indicators to quantify the potential supply of the AESs. Function Indicators (FIs) that relate to an ecosystem function are currently being built using Modelbuilder in Arcgis 10.3. Nine FIs were selected following five criteria: C1 Simple and easily measured, C2 Data availability, C3 Relate at the appropriate scale (temporal and spatial), C4 Comparable, repeatable and defensible between sites and times, and C5 Present a relationships cause/effect between AESs and forest activities. So far, the values for four of the nine FIs have been calculated for each sub-watershed. Correlation analysis is then used to determine the strength of the relationships between FIs. The most responsive and independent FIs will be combined into a single index to represent the potential supply of AESs. The preliminary identification and correlation analysis of four selected Function Indicators is summarized in Figure 2.
Ongoing work is focused on calculating the remaining five FIs, developing a weighting system for the AES index, and evaluation of the method. Five maps will be created between 1994 to 2018 that show the cumulative effects of 5 years of forest activities.
Figure 1: Harry’s River Watershed delineation. The challenge was to define the lake outlets as sub-watershed outlets to facilitate a potential field data acquisition on water quality. For the calculation of the FIs, the three big lakes were extracted from the watershed layer and FIs were calculated on the terrestrial part only.
Figure 2: Pearson’s correlation analysis. Forest FI corresponds to the percent of forest cover, Crossing FI corresponds to the number of forestry roads cross a river reach, Wet FI and Imp FI correspond to the percent of wetland and impervious surface respectively. The preliminary results show a high correlation between the forest cover and wetland cover.
Wall-to-wall map of individual tree species at JD Irving’s Black Brook Forest, NB
As part of the AWARE project and in collaboration with JD Irving, a wall-to-wall map of individual tree species at the Black Brook Forest was completed in April of 2017. Led by Benoît St-Onge, a professor at UQAM and AWARE investigator, Rachel Perron (M.Sc. student, Q22), with assistance from Jean-François Prieur (Ph.D. student, Q21) developed special methods and software to analyse a huge airborne lidar dataset for predicting the species of every single tree visible from the air. This is no small feat as the Black Brook forest occupies 2090 km2 and is populated by more than 145 million trees. Despite there being a good number of scientific studies on the identification of trees from liDAR on small patches of forests, there is no evidence that such a large number of trees had ever characterized in an entirely automated manner.
The process involved the gathering of reference tree crown samples in the field from existing sample plots and plantation maps, and through expert photo-interpretation to train a machine learning classifier. In parallel, the airborne liDAR tiles were converted to canopy height models and single tree crowns were automatically delineated using the SEGMA application developed at Benoît St-Onge’s lab, resulting in tens of millions of polygons. The lidar point clouds were then extracted for each crown polygon and stored in individual files. Classification features (“metrics”) were computed for each crown based on the characteristics of the individual point cloud and stored in a PostGIS database. Features included three-dimensional metrics (e.g., the average slope of the crown profile, the crown “porosity” to laser pulses, etc.) and intensity metrics (e.g., average and standard deviation of intensity per crown, etc.). The reference crown samples were used to train random forest classifications based on all these features. Several classification levels were employed to classify trees as deciduous/coniferous, at the genus, and species levels. Specific classification models were developed to separate spruces from balsam firs. Based on a training set composed of more than 1 500 crowns, the accuracies were 49% at species level, 58% at genus level, and 93% for deciduous/coniferous separation.
Once validated, the classification models were applied to all the individual crowns. This resulted in more than 2000 map tiles, each 1 km x 1 km and containing on average 75 000 processed crowns, i.e. polygons for which the tree position, size (height, crown area, etc.), and predicted species at different levels are recorded (see figure below). The sheer volume of the input and output data represented a challenge in itself. Simple operations such as clipping liDAR points with the crown polygons just could not be done in a reasonable amount of time with most existing GIS applications. To get around this lack of functionality, Rachel Perron developed specific software approaches for optimizing each process and digesting the enormous amount of data. This required significant effort and time, but resulted in a unique, high capacity set of tools.
In addition to the error estimation performed on the reference crowns, JD Irving staff are currently checking the output species information in the field. The inspection reveals that the classification performed as expected in some situations, but may be less reliable in others. This type of “real world” verification is rarely reported in the scientific literature and can only be done if a wall-to-wall map was generated. What can be learned from this is, among other issues, that the selection of sample crowns for such a large territory may represent a greater challenge than expected. For example, for any given species, trees of different size class might have to be sampled in a systematic manner. Such a posteriori validation provides a clearer view of new and important research questions. In addition to sampling strategies, future research will consist of including contextual information such as drainage and soil type, as well as additional remote sensing data in the form of multispectral airborne images to improve the accuracy of species identification.
Figure – Individual tree crown polygons coloured by predicted species (overlaid on a color infrared airborne image).
This summer, IUFRO will hold a conference in Vancouver from June 12-16th (www.IUFROdiv5-2017.ca). The week after, from June 20-22nd, Earth Observation Summit will be happening at UQAM. AWARE is planning to present at both these events next summer.
AWARE Year 2 Annual General Meeting
AWARE’s 2nd AGM will be held from May 16-18th, 2017 at the Madawaska Historical Museum in Edmundston, New Brunswick. Please see our news page for registration and the preliminary agenda.
If you are making your own travel arrangements for the AGM, a block of rooms has been set aside for AGM attendees at the Best Western, Edmundston. Book Best Western