Identifying Drivers and Mechanisms of Climate-Related Tree Mortality
(2025)- Abstract
- Eastern New South Wales (NSW), Australia, is a region prone to frequent and intense droughts. An unprecedented episode of high temperatures and low rainfall in 2017–2019 culminating in the 2019-2020 summer triggered massive canopy dieback of eucalypt forests throughout the region. Subsequent high rainfall in 2020 offered a unique and rare opportunity to study the process of vegetation recovery following severe drought impact. Pinpointing the environmental triggers and physiological responses that drive tree stress and recovery needs to account for complex ecological interactions unfolding over time particularly in heterogeneous environments. This research addresses these complexities by quantitatively linking ground-based observations of... (More)
- Eastern New South Wales (NSW), Australia, is a region prone to frequent and intense droughts. An unprecedented episode of high temperatures and low rainfall in 2017–2019 culminating in the 2019-2020 summer triggered massive canopy dieback of eucalypt forests throughout the region. Subsequent high rainfall in 2020 offered a unique and rare opportunity to study the process of vegetation recovery following severe drought impact. Pinpointing the environmental triggers and physiological responses that drive tree stress and recovery needs to account for complex ecological interactions unfolding over time particularly in heterogeneous environments. This research addresses these complexities by quantitatively linking ground-based observations of eucalypt drought-stress with spectral observations from remote sensing, employing advanced machine learning (ML) and deep learning (DL) methods, and comparing outcomes with a process-based ecosystem model to unravel drought and recovery dynamics and underlying ecological response mechanisms in eucalypt ecosystems. This research comprises four interconnected studies. Subproject-1 established relationships between plot-based canopy health observations and satellite-derived spectral indices, scaled these to map regional drought impacts, and validated results against a citizen-science dataset, confirming severe drought stress in approximately 13% of eastern NSW forests. This approach demonstrated that Normalized Burn Ratio (NBR) index is a reliable tool for establishing long-term baselines to monitor drought responses. In Subproject-2, Artificial Neural Networks (ANNs) were employed to link field-validated z-scores of NBR index to dynamic and static spatio-temporal environmental variables hypothesised to represent key drought-stress and recovery drivers. Findings showed lagged, cumulative climatic stresses and landscape factors such as wetness index, aspect, and soil properties governed drought impacts and subsequent recovery, and that these effects vary across bioregions according to local environmental conditions. Subproject-3, introduced a hybrid Convolutional Neural Network – Long Short-Term Memory (CNN-LSTM) model to link the same NBR z-score with dynamic and static spatio-temporal environmental features, finding that in dry to mid-humid regions resilience hinges on climatic stressors (e.g., PET and rainfall variability) and landscape features (soil and topography), whereas in wet-humid areas post-drought recovery is driven by interactions between climate and terrain. This framework outperformed conventional ML by better capturing complex spatiotemporal patterns. Subproject-4 aimed to elucidate the mechanistic drivers of drought stress by integrating ecosystem model simulations with machine-learning inferences and the field-calibrated spectral index from previous sub-projects. This unified approach revealed hydraulic-impairment signatures, rapid upper-soil moisture loss and carbon depletion at high-severity sites, and deeper soil-moisture buffering at moderate-severity locations, while exposing the model’s shortcomings (static trait parameterization and absence of root plasticity). Based on the findings, it is hypothesized that incorporating hydraulics traits may enhance the model’s predictive accuracy under severe drought. These interconnected studies provide scalable methodologies for drought assessments that could empower policymakers to prioritize vulnerable ecosystems and allocate resources toward forest resilience-building interventions. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/876f10ed-fcbf-4cea-9702-73e5f4e92c27
- author
- Arampola, Arampola Mudiyanselage Nuwanthi Sashiprabha LU
- supervisor
-
- Ali Mansourian LU
- Benjamin Smith LU
- Stefan Olin LU
- Pengxiang Zhao LU
- opponent
-
- Professor Stoy, Paul, University of Wisconsin
- organization
- publishing date
- 2025-08-01
- type
- Thesis
- publication status
- published
- subject
- keywords
- Drought, Remote Sensing, Machine Learning, Deep Learning, Ecosystem Modelling, Pixel Based, Grid Based, Canopy and Woodland
- pages
- 295 pages
- publisher
- Lund University
- defense location
- Världen, Geocentrum I, Sölvegatan 10, Lund.
- defense date
- 2025-10-10 10:00:00
- ISBN
- 978-91-89187-64-1
- 978-91-89187-63-4
- language
- English
- LU publication?
- yes
- id
- 876f10ed-fcbf-4cea-9702-73e5f4e92c27
- date added to LUP
- 2025-08-28 14:26:45
- date last changed
- 2025-09-15 10:39:21
@phdthesis{876f10ed-fcbf-4cea-9702-73e5f4e92c27,
abstract = {{Eastern New South Wales (NSW), Australia, is a region prone to frequent and intense droughts. An unprecedented episode of high temperatures and low rainfall in 2017–2019 culminating in the 2019-2020 summer triggered massive canopy dieback of eucalypt forests throughout the region. Subsequent high rainfall in 2020 offered a unique and rare opportunity to study the process of vegetation recovery following severe drought impact. Pinpointing the environmental triggers and physiological responses that drive tree stress and recovery needs to account for complex ecological interactions unfolding over time particularly in heterogeneous environments. This research addresses these complexities by quantitatively linking ground-based observations of eucalypt drought-stress with spectral observations from remote sensing, employing advanced machine learning (ML) and deep learning (DL) methods, and comparing outcomes with a process-based ecosystem model to unravel drought and recovery dynamics and underlying ecological response mechanisms in eucalypt ecosystems. This research comprises four interconnected studies. Subproject-1 established relationships between plot-based canopy health observations and satellite-derived spectral indices, scaled these to map regional drought impacts, and validated results against a citizen-science dataset, confirming severe drought stress in approximately 13% of eastern NSW forests. This approach demonstrated that Normalized Burn Ratio (NBR) index is a reliable tool for establishing long-term baselines to monitor drought responses. In Subproject-2, Artificial Neural Networks (ANNs) were employed to link field-validated z-scores of NBR index to dynamic and static spatio-temporal environmental variables hypothesised to represent key drought-stress and recovery drivers. Findings showed lagged, cumulative climatic stresses and landscape factors such as wetness index, aspect, and soil properties governed drought impacts and subsequent recovery, and that these effects vary across bioregions according to local environmental conditions. Subproject-3, introduced a hybrid Convolutional Neural Network – Long Short-Term Memory (CNN-LSTM) model to link the same NBR z-score with dynamic and static spatio-temporal environmental features, finding that in dry to mid-humid regions resilience hinges on climatic stressors (e.g., PET and rainfall variability) and landscape features (soil and topography), whereas in wet-humid areas post-drought recovery is driven by interactions between climate and terrain. This framework outperformed conventional ML by better capturing complex spatiotemporal patterns. Subproject-4 aimed to elucidate the mechanistic drivers of drought stress by integrating ecosystem model simulations with machine-learning inferences and the field-calibrated spectral index from previous sub-projects. This unified approach revealed hydraulic-impairment signatures, rapid upper-soil moisture loss and carbon depletion at high-severity sites, and deeper soil-moisture buffering at moderate-severity locations, while exposing the model’s shortcomings (static trait parameterization and absence of root plasticity). Based on the findings, it is hypothesized that incorporating hydraulics traits may enhance the model’s predictive accuracy under severe drought. These interconnected studies provide scalable methodologies for drought assessments that could empower policymakers to prioritize vulnerable ecosystems and allocate resources toward forest resilience-building interventions.}},
author = {{Arampola, Arampola Mudiyanselage Nuwanthi Sashiprabha}},
isbn = {{978-91-89187-64-1}},
keywords = {{Drought; Remote Sensing; Machine Learning; Deep Learning; Ecosystem Modelling; Pixel Based; Grid Based; Canopy and Woodland}},
language = {{eng}},
month = {{08}},
publisher = {{Lund University}},
school = {{Lund University}},
title = {{Identifying Drivers and Mechanisms of Climate-Related Tree Mortality}},
year = {{2025}},
}