Skip to main content
Overview

Mr. Matthew Wiecek

Research Postgraduate (PhD)

M.A.


Affiliations
AffiliationRoom numberTelephone
Research Postgraduate (PhD) in the Department of GeographyW601 

Biography

My Research

For my PhD dissertation research, I am developing improved methods of using satellite images to monitor deforestation in South America, in collaboration with the Government of Guyana. For 12 years, Guyana has been working hard to minimize deforestation in its tropical rainforest. They get financial support from the Government of Norway, which is tied to the amount of deforestation that Guyana prevents. To measure deforestation, they have a web application that uses satellite images and machine learning to identify deforestation events with good but not perfect accuracy. As a PhD student, I’m working in collaboration with Guyana to develop an improved method of automatically detecting deforestation events in real time. Deforestation is a major factor in environmental degradation, economic development, local quality of life and climate change, and my research will give national policymakers, local governments and climate change forecasters data that is more accurate, more precise and more detailed.

The policy context of my research is the REDD+ program. REDD+ stands for Reducing Emissions from Deforestation and Forest Degradation, and it is a United Nations-developed framework for developing agreements where a developed country provides financial support to a developing country in support of deforestation mitigation. A key part of a REDD+ program is a Monitoring, Verification and Reporting Systems (MRVS). An MRVS spells out how the developing country will measure deforestation, conduct audits and accuracy assessments, and report the amount of deforestation that they prevented or reversed in each year. The amount of funding provided is tied to the amount of deforestation that is prevented, hence the critical importance of accurate and detailed deforestation mapping.

The first REDD+ agreement was signed by Norway and Guyana in 2009. This program has grown considerably over the last 12 years. In the beginning, they only mapped change between forest and non-forest land cover types. Now, they map expansion and contraction of forested areas, changes in carbon stocks and biomass, detailed land cover/land use change mapping, collaboration with local communities, and more. As of 2016, they have also made scientific journal publications an objective of their REDD+ program.

 

The government agency responsible for the Guyana-Norway REDD+ agreement is the Guyana Forestry Commission. They have their in-house policymakers and technicians, and external contractors, who include Indufor Asia-Pacific and Prof. Danny Donoghue of Durham University. Danny has been a consultant in their accuracy assessment and auditing work for several years, and Indufor provides technical assistance in monitoring deforestation. To this end, Indufor has developed a web application that uses radar satellite images and machine learning to automatically identify deforestation events. Their accuracy and detail are good but have room for improvement.

As a PhD student, I’m doing a detailed analysis of the characteristics and behaviour of a variety of datasets and analysis methods. The datasets I’m working with include:

  • Visible light and infrared images from satellites such as Landsat, Sentinel-2, and Planetscope. The slice of the infrared spectrum called Near Infrared is very strongly reflected by chlorophyll, which makes it a very useful measure of vegetation vitality. The range of wavelengths called Shortwave Infrared is reflected by moisture in vegetation. Landsat also measures Thermal Infrared, which is useful in measuring temperature.
  • Radar satellites such as Sentinel-1. Radar is very good at measuring the structure and moisture content of a forest.
  • In addition to examining each dataset in isolation, I’m also looking at possibilities for integrating them to one workflow that combines the unique information that’s found in one data type but not another. Combining information from varied sources has the potential to reveal details not visible in either one alone.

Automatic identification of deforestation events will be done using image classification. Classifying images from different times, and looking at the difference, is how change is detected. In classification, the input is each pixel of an image, and the output is two values: forest, and non-forest. The goal is to set up a web application that takes a satellite image, and uses it to calculate where forest and non-forest are in the image.

There are two ways to do this. One way is to give the computer a model that tells it how to identify forest and non-forest in an image. This requires you to know what that model is in advance. The machine learning way is to give the computer the satellite image, and the correct answer, and have the computer figure out what the model is for you. If the computer did a good job and the model is accurate, then you will be able to give the computer a new image that it has not seen before, and be confident that the classification is accurate.

To be sure that the model is good, you need to test it. This is a two part process with five datasets:

  1. Training. In this stage, the computer is given two datasets: the training data and the correct result. The computer fits a model to the data.
  2. Testing. In this stage, there are three datasets: the testing data, the predictions, and the correct result.
    1. The testing data is a different satellite image that the computer has not seen before. The testing image is given to the model fitted above, and the model makes predictions about which parts of the image are forest and non-forest.
    2. The predictions are then compared with the correct result. The difference between the two is the accuracy of the predictions. If the difference is small, then the accuracy is high and the model is good. If the difference is great, then the accuracy is low and the model is bad.

The model is a machine learning algorithm. Each algorithm is a different mathematical/statistical process for classifying data, and different algorithms work well with different kinds of data. I will test a variety of algorithms on the data, and record detailed information about how they behave and what exactly they produce. This will enable me to understand

I’m working to provide something for three stakeholders:

  1. I am working to produce original research that I can then submit as a dissertation, and get a PhD.
  2. I am going to work to ensure that my published research meets the Guyana Forestry Commission’s objective of producing scientific journal publications as a result of the REDD+ program.
  3. I am working to give Indufor recommendations on how to improve their process.

After this research is done, I’m interested in collaborating with other researchers who benefit from highly accurate and detailed forest change data. This includes climate change forecasting, forest ecology, the connection between deforestation and infectious disease outbreaks, and more.

Research interests

  • Spatial data analysis using geographic information systems
  • Remote sensing, especially sensor fusion, image classification, and time series analysis
  • Environmental archaeology, especially human-climate interaction before the Industrial Revolution
  • Reforestation and agroforestry science and policy, including land management practices of past cultures and present non-Western cultures

Publications

Conference Paper

  • Peter Zahradka, Danielle Perera, Angela Wilson, Matthew Wiecek, Rhonda Bell & Carla G Taylor (2018), Improvements in arterial stiffness due to bean and pea consumption are determined by metabolic state, The FASEB Journal 31: Experimental Biology. Chicago, Federation of American Societies for Experimental Biology, 963.5.

Journal Article