Skip to main content Skip to local navigation

Learning post-disturbance boreal recovery trajectories for backward prediction

Learning post-disturbance boreal recovery trajectories for backward prediction

by Philip Lynch

Philip Lynch

Stand-replacing disturbances (i.e., harvesting, forest fires) in Canadian forests and recovery cycles cause highly dynamic landscapes, demanding continuous monitoring to characterize and observe ecological modification over time. Monitoring large-scale post-disturbance recovery by satellite remote sensing is a major research area in Canada. My research seeks to answer the question: What improvement do shortwave- and thermal-infrared-based vegetation indices offer for discerning boreal disturbance types, intensities, and recoveries, relative to traditional visible- and near-infrared-based indices?

Four independent investigations are proposed to examine specific aspects of this query, but that collectively contribute to a greater understanding of the whole system. First, a comprehensive comparison of vegetation indices will be undertaken to identify their structural biases, their sensitivities, and best practices for their use. Second, an assessment of time-series and spatial variability of the most applicable vegetation indices identified will be conducted to test for site influences other than vegetation condition (e.g., terrain, soil, ecoregion). Third, boreal vegetation recovery trajectories will be learned by scanning per-pixel vegetation index values through time, producing the largest ever dataset of recovery at the pixel-level. These recovery trajectories will be clustered into like forms and correlations tested among competing explanatory cases. Fourth, the learned recovery trajectories will be used to back-project disturbance origins where data is missing, occluded, or before the 1972 availability of satellite imagery.

The Ministry of Northern Development, Mines, Natural Resources and Forestry (NDMNRF) is responsible for monitoring public Crown forests, including requiring compliance with forest management plans filed by forest harvesting companies, in addition to taking forest inventories on non-licensed and private lands. While many survey methods exist to monitor the environment, field methods by themselves cannot reach the scale of Ontario's entire boreal zone on a routine basis, nor can field surveys be performed in areas difficult to access by humans, such as large forests without roads. Satellite remote sensing is an effective method for conducting geographic surveys over vast extents and long periods, with the potential to advance our knowledge of environmental processes over the time-series of collected data with the added advantage of being, typically depending on how high the resolution of a sensor, a free data source.

Satellite remote sensing technology is valuable for effectively detecting changes in forest vegetation conditions at many scales. A challenge exists where a single remotely sensed cell's reflectance might contain signals from multiple features, depending on the spatial resolution. Methods for enhancing each pixel must sometimes be leveraged to extract as much information as possible from satellite data. The most common image enhancement techniques are called spectral enhancement indices. Spectral indices focus on mathematically magnifying characteristic spectral properties of target features, often by taking the ratio of contrasting spectral responses from key wavelength ranges within the electromagnetic spectrum, to enhance and classify target features within a gradient of continuous values. Among the many different types of indices available for monitoring Earth features, vegetation indices are most commonly utilized. Remote sensing vegetation indices analyses are conducted primarily by leveraging reflectance in the visible (VIS) blue (450 – 495 nm), green (495 – 570 nm), red (620 – 750 nm), and invisible near-infrared (NIR) (850 nm – 1,700 nm) wavelengths. Specifically, incorporating the NIR wavelength has been proven advantageous for magnifying specific vegetation health components such as moisture, pigments, carbohydrates, protein, and aromatics.

However, it has been shown that vegetation indices incorporating band combinations within the invisible shortwave-infrared (SWIR) (1,500 nm – 5,000 nm) and thermal-infrared (TIR) (8,000 nm – 12,500 nm) wavelengths exhibit a stronger correlation with biophysical properties of vegetation (e.g., canopy structure and leaf area) compared to indices incorporating only the VIS or NIR spectral regions. Most specifically, Boyd et al. (1996) report that, compared to vegetation indices incorporating only VIS and NIR bands, they boast significantly higher correlation with plant biophysical properties, specifically stand age for the Amazon rainforest. To automatically assess vegetative regeneration stages of forests, according to how long they have aged after deforestation based on a low-high range of values derived from imagery, the infrared vegetation index (VIIR), incorporating SWIR and TIR given, was given as:


Comparison of 2007 and 2020 VIIR derived from Landsat 7 data, depicting regeneration from the 2003 Okanagan Mountain Park Fire in British Columbia. The fire burned roughly 250 sq km, an area containing a majority of the park's trees.

It is assumed that VIIR will discriminate greater variability (heterogeneity) among landscape features and better discriminate Canadian boreal landscape changes compared to previous traditional indices.

The study area for this research will be the 30 million ha area of managed forest defined by the NDMNRF based on disturbance frequency. Image archives provided by the United States Geological Survey for the moderate-scale Landsat satellite constellation will be utilized in cross-reference with Canadian data records, including the digital elevation model, climate, land cover, and point-based disturbance inventories. Projected outcomes and significance include guidelines for selecting the best vegetation indices for an application and a clear understanding of limits and application domains for select vegetation indices. Additionally, we will use statistical recovery curves derived from vegetation indices to predict and back-cast regeneration unobservable in remote sensing archives. For instance, anything pre-dating 1972, areas omitted from the Landsat 7 ETM+ image swath due to the scan line error that began on 31 May 2003, and any data between the failure of Landsat 5 TM on 18 November 2011 and the first acquisitions made by Landsat 8 OLI on 13 March 2013.

Philip Lynch is a PhD candidate in Geography at the Faculty of Environmental and Urban Change under the supervision of Prof. Tarmo Remmel. They recently presented their work on Satellite remote sensing post-fire regeneration assessments at the Geographic Information Systems (GIS) Day in November 2021.


Boyd, D.S., Foody, G.M., Curran, P.J., Lucas, R.M., Honzak, M., 1996. An Assessment of Radiance in Landsat TM Middle and Thermal Infrared Wavebands for the Detection of Tropical Forest Regeneration. International Journal of Remote Sensing 17, 249–261.