Drought Predictive Modelling and Ecosystem analysis Using ML

NAISS 2024/5-52


NAISS Medium Compute

Principal Investigator:

Ali Mansourian


Lunds universitet

Start Date:


End Date:


Primary Classification:

10507: Physical Geography

Secondary Classification:

10299: Other Computer and Information Science

Tertiary Classification:

10611: Ecology




Tree stress and mortality resulting from climate variability and change are significant global concerns, particularly in the context of increasing extreme weather events like droughts and heatwaves. The Eastern Australian region experienced an exceptional event from 2017 to 2020, marked by a prolonged drought period with extensive browning observed in Eucalyptus forests. A subsequent period of regular rainfall from early 2020 created favorable conditions for forest canopy recovery. This unique sequence of events presents an extraordinary opportunity to investigate the underlying drivers and mechanisms behind tree dieback and subsequent recovery. Understanding the precise factors that trigger vitality loss and the mechanisms associated with tree stress under drought is complex due to the multifaceted and time-dependent nature of responses. Furthermore, the relative importance of different drivers varies across species, bioregions, and durations of exposure to extreme conditions. Field measurement campaigns as well as citizen science initiatives such as "Dead tree detective" provide invaluable ground-based observations to assist in identifying areas affected by tree browning and the responsible processes and responses. Remote sensing using satellite observations provides a way to extrapolate insights from ground-based observations in space and time. This project aims to employ spatial-temporal analysis and machine learning, to link field observations, remote sensing and other environmental datasets to investigate the impacts of drought on tree health in Eastern Australian woodlands and forests, and to identify the underlying causes of tree health decline. Machine learning techniques will be utilized to quantitatively analyse spatial information from diverse sources, with results compared to predictions from a process-based vegetation model for mechanistic insight. In summary, the research demonstrates how the complementary strengths of observations and models may be brought together to enhance mechanistic understanding of canopy health variations and develop models that connect local-scale observations to broader regional landscapes while considering underlying causes.