PhD Defence Sam Hislop

using earth observation satellites to explore forest dynamics across large areas

Sam Hislop is a PhD student in the department of Natural Resources (NRS). His supervisors are prof.dr. A.K. Skidmore from the faculty of Geo-Information Science and Earth Observation (ITC) and prof.dr. S. Jones from RMIT University.

A third of the land on earth is covered by forests. Forests provide valuable resources and essential ecosystem services, including filtering air and water, harbouring biodiversity and managing the carbon cycle. Regular monitoring and reporting across various indicators is necessary to manage forests sustainably. Due to the vastness of forests, satellite Earth observation is one of the most practical and cost-effective ways to monitor forests. The regular and consistent measurements provided from space enable time series analysis, which can reveal trends over time. The temporal, spatial and radiometric depth of the Landsat archive, which extends back to 1972 in some cases, is one of the most useful resources for monitoring forest dynamics across large areas.

Analysing forest disturbance and recovery trends using Landsat has recently become widespread, particularly since the opening of the image archive in 2008. However, deriving useful information from the data is challenging on many fronts, including overcoming cloud-cover, differentiating true changes from noise and relating spectral measurements to meaningful outputs. In addition, large data volumes create hurdles for processing and storage. This study presents new techniques for exploiting the Landsat archive in relation to monitoring and measuring forest disturbance and recovery across large areas. Landsat data were processed through a series of steps, analysed in time series, and combined with other data sources to produce mapped outputs and statistical summaries, which can be interpreted by non-experts. The spatial extent of the analysis expands across multiple scales – from local and regional to global (temperate and boreal forests).

Firstly, eight Landsat spectral indices were assessed to determine their sensitivity to forest disturbance (caused by wildfire) and recovery in southeast Australian forests. Results indicated that indices making use of the shortwave infrared wavelengths were more reliable indicators of forest disturbance and recovery than indices using only the red and near-infrared wavelengths. Following this exploratory analysis, three indices and two change detection algorithms were evaluated in terms of their ability to detect forest disturbance. Results showed that the LandTrendr algorithm with the Normalised Burn Ratio (NBR) was the most accurate single algorithm/index combination (overall error 21%). However, results were greatly improved by using an ensemble approach. A Random Forests model combining several Landsat-derived metrics with multiple indices, trained with human interpreted reference data, had an overall error of 7%. A notable finding was that priming the training data with confusing cases (commission errors from the change detection algorithms) led to increased accuracy.

One Random Forests model was used to create annual forest disturbance maps (1989-2017) across the state of Victoria, Australia. These maps, in conjunction with each pixel’s temporal trajectory, were used to extract metrics for spectral disturbance magnitude and recovery length across 2 million ha of burned forest in southeast Australia. The association between disturbance magnitude and forest recovery length, as measured spectrally, was then explored. A novel patch-based technique was used to isolate the disturbance-recovery relationship from confounding factors such as climate, elevation and soil type. The results showed statistically significant differences across bioregions and forest types. The patch-based method demonstrated how Landsat time series can be harnessed to explore ecological changes.

The methods developed above were then employed over a much larger area, to investigate trends in fire disturbance and forest recovery in temperate and boreal forests worldwide. This work used both MODIS and Landsat data, through the Google Earth Engine platform, to look at trends in burned area, fire severity and forest recovery across almost 2 billion ha of forests, over the last 18 years. Burned area results showed significant increasing trends in two cases: coniferous forests in Canada and Mediterranean forests in Chile. A significant decreasing trend was found in temperate mixed forests in China. An assessment of fire severity, as measured by Landsat spectral change, highlighted possible trends in a few cases; most notably, the Russian taiga, where increasing severity was observed. An analysis of forest recovery, based on Landsat time series, indicated recovery times were accelerating in many regions. However, given the relatively short time-period analysed, these results should be interpreted with caution.

The results presented in this thesis demonstrate the power of Earth observation satellites in monitoring forests at the landscape scale. Although forests are complex systems that are influenced by a myriad of factors, the regular and consistent measurements provided by satellites can be analysed in time series to provide inter-comparable results across large areas. This can broaden our understanding of the dynamic nature of forests, and in doing so, help progress towards their sustainable management.