Using satellite data to improve mapping of stem density and forest carbon for sustainable forest management

Forests are important due to the many ecosystem services they provide. The most important one being climate protection because they act as a carbon sink: Sustainable forest management and use of the natural resource timber helps in reducing carbon dioxide. As forests remove carbon dioxide from the air they contribute to reducing green house gas in the atmosphere. Timber stores carbon dioxide and by replacing other materials with timber this effect is enhanced. In addition forests provide a source of livelihood, recreation, wildlife habitat, keep the climate in balance, filter pollutants and fine particles, help maintain the hydrological balance, protect from soil erosion and avalanches and secure drinking water.

This project is an exciting opportunity to get involved in interdisciplinary research in statistics, data and forest science, to improve mapping of forest carbon for monitoring forest properties, and to help with mitigation of climate change effects. It will equip you with important skills in remote sensing, statistical modelling and machine learning, as well as data science skills, with hands on opportunities to learn about forest science. You will collect field data in Scotland in partnership with forest managers, and spend 3 months on a placement within your CASE Partner (Forest Research).

Concentrating initially on European temperate forests we will explore the following questions:
1. Review and validation of current state of the art methods for mapping forest carbon and stand density. Model validation: Validation of how well the models predict forest variables, using the field data as ground truth.
2. Statistical Design: What is the optimal combination of field and remotely sensed data, i.e. optical allocation of resources for mapping forest carbon. Does this optimum depend on the forest type being mapped (i.e. even-age monoculture stands and mixed species/age woodlands).
3. Model development and validation: Improving statistical models for mapping; state of the art method use of the shelf models (multiple regression, k-nearest neighbours, support vector regression, random forest algorithms) . The developed model will likely involve fusion of data obtained at point (field plot tree level data) and areal resolutions (pixels from satellite and lidar data). Possible avenues for this are machine learning methods for merging different data sources (see e.g. Baez-Villanueva et al, 2020) or generalized spatial fusion models implemented using a Baysian hierarchical framework (see e.g. Wang et al, 2017, Cameletti et al, 2019). Model validation using the field data as ground truth.
4. How to solve the signal saturation problem. Does incorporating LiDAR and field plot data help with this?
5. How well do the methods translate to other types of forest (temperate or tropical)?

The tools and data to be used include:
1. Software: R, INLA, Python, Stan
2. Datasets:
o Scotland level data from a combination of managed single-age stands (mostly of Sitka Spruce trees) and mixed native woodlands, based on field plots from the CASE partner and open access aircraft-derived LiDAR data.
o Forest inventory and LiDAR datasets for the Rioja region of Spain and the whole of Denmark, from their National Forest Inventories, including hundreds of thousands of tree measurements from about 1700 plots used in Joshi (2017).
o Using the expertise of Tim Baker, open access forest plots from unmanaged forests across the tropics
o In all cases we also will be using the corresponding data from L-band synthetic aperture radar (SAR) from the ALOS-2 satellite, C-band SAR data from the Sentinel-1 satellite, and optical data from the Sentinel-2 satellite.