Supervisor Team: Kyle Dexter (Primary; University of Edinburgh); Renata Françoso (Universidade Federal de Lavras, Brazil); Tim Baker and Oliver Phillips (University of Leeds)
Case Partner: Space Intelligence (https://www.space-intelligence.com/), via Murray Collins
Summary: The global extent and distribution of tropical dry forests versus savannas is poorly known. The project will use Artificial Intelligence applied to new data from recently launched satellites to map dry forests versus savannas across the globe. The student will then estimate their carbon storage and its change over the last 15 years.
Background and Motivation: Earth observation approaches that aim to map the distribution and carbon storage of forests are generally successful for tropical moist forests, but have been less so for tropical dry forests. Tropical dry forests are shorter in stature and may not have a closed canopy, and are thus frequently confused with wooded savannas in remotely sensed imagery and products (Beuchle et al. 2015). Yet, knowing the distribution and relative extent of tropical dry forests versus savannas is imperative to understanding the carbon cycle of the dry tropics and to develop appropriate management strategies for tropical vegetation (Dexter et al. 2018). Tropical dry forests store more above-ground carbon than savannas, while fire is damaging to tropical dry forests and their carbon storage ability. In contrast, fire is necessary to the ecosystem function and biodiversity of savannas. This project will improve ecosystem management in the dry tropics and quantify the role of the dry tropics in global carbon cycles.
Key Research Questions:
Q1) What machine learning/AI techniques and remotely sensed information (e.g. phenology, fire occurrence, canopy height) best predicts tropical dry forest versus savanna occurrence?
Q2) Where do tropical dry forests versus savannas occur across the globe, and what are the major environmental and anthropogenic drivers of their distribution?
Q3) How much carbon is stored in dry forests versus savannas, and how has this changed from the mid-2000s to 2020, due either to land conversion or changing tree density?
Methodology: Cloud-based analysis of new remote sensing products, using the latest artificial intelligence (AI) and machine learning (ML) algorithms, will allow us to better distinguish tropical dry forests from savannas, and to successfully map the distribution of dry forests across the tropics (Q1, Q2, Q3). This project will make use of the extensive freely-available data sets from the ESA Copernicus missions: Sentinel 1A/B for C-band SAR; and Sentinel 2A/B for Multispectral optical sensing. These will be supplemented by Landsat data. Lidar sample data will be obtained from the GEDI mission. (Q1, Q2). The data products will be used to map maximum canopy height and percent canopy cover across the dry tropics, which will in turn allow estimation of above-ground biomass (AGB) and its change from the mid-2000s to the present (Q3). The project will depend on ground-truthed mapping of tropical dry forests versus savannas (Miranda et al. 2018) (Q1), and a dataset of canopy height, tree density and AGB from >2000 plots in tropical dry forests and savannas being assembled by the supervisors as part of a recently funded NERC Large Grant (Q3). AI and ML approaches will be used to build statistical models that relate ground-truthed information to EO data and enable wall-to-wall mapping (Q1,Q2).