Tracking tree demography and forest dynamics at scale using remote sensing

Abstract

Introduction

Forest ecosystems face growing pressure on multiple fronts, from increasingly frequent and severe droughts and heatwaves, larger and more intense wildfires and storms, novel pests and pathogens, and human-driven degradation (Senf et al., 2018; Canadell et al., 2021; Hammond et al., 2022; Turner & Seidl, 2023). Understanding how trees are responding to these novel disturbance regimes is critical if we are to forecast how forest dynamics will change over the coming century, and what implications this will have for biodiversity and carbon storage in these ecosystems (McDowell et al., 2020; Turner & Seidl, 2023). To achieve this, we need demographic information that allow us to infer and model changes in population dynamics at scale – data that capture how rates of tree growth, mortality and recruitment vary across both space and time (Coomes et al., 2014; Fisher et al., 2018; Kunstler et al., 2021; Needham et al., 2022b; Zuidema & van der Sleen, 2022). Ecologists have traditionally relied on networks of permanent field plots to estimate these demographic rates (Lines et al., 2010; Ruiz-Benito et al., 2013; Kunstler et al., 2021; Needham et al., 2022b; Piponiot et al., 2022). However, while plot networks remain the gold standard to characterise community-level dynamics, they have some inherent limitations when it comes to capturing variation in demographic rates across landscapes. Field surveys are incredibly labour-intensive, meaning that most forest plots are small (0.1–1 ha) and cumulatively only cover a tiny fraction of the total forest area (< 0.01% even in best-case scenarios; Yu et al., 2022; Holcomb et al., 2023). This makes it challenging to understand how demographic rates vary across environmentally heterogeneous landscapes and in response to large, infrequent disturbances.

Remote sensing offers an intuitive solution to this challenge of tracking large numbers of trees across broad spatial scales (Stovall et al., 2019; Brandt et al., 2020; Ma et al., 2023). In particular, technologies such as airborne laser scanning (ALS, or LiDAR) can be used to build highly accurate and detailed 3D models of both the forest canopy and the underlying terrain (≤ 1-m resolution) that span thousands of hectares (Jucker, 2022; Lines et al., 2022). Unsurprisingly, ALS has become an integral tool for large-area mapping of forest structure and biomass, and there is now a growing interest in using repeat ALS acquisitions to quantify forest dynamics at scale (Asner & Mascaro, 2014; Dalponte et al., 2019; Cushman et al., 2021; Dalagnol et al., 2021; Nunes et al., 2021; Jucker et al., 2023). To date, almost all this work has focussed on characterising dynamic processes occurring at a canopy level, such as those associated with the formation, expansion and closure of gaps (Wedeux et al., 2020; Cushman et al., 2021; Dalagnol et al., 2021; Nunes et al., 2021; Choi et al., 2023). However, in parallel, there has also been a push towards developing computational tools to detect individual tree crowns using ALS (Dalponte & Coomes, 2016; Ferraz et al., 2016; Cao et al., 2023). Benchmarking efforts have shown that algorithms can accurately segment and measure the crown dimensions of canopy-dominant trees (Wang et al., 2016; Cao et al., 2023), particularly when applied to high-resolution ALS data acquired in open canopy forests. As repeat ALS data become increasingly available, individual-based methods provide a unique opportunity to characterise how rates of tree growth and mortality vary across size-structured populations (Piponiot et al., 2022; Brandt et al., 2024) and whole landscapes (Duncanson & Dubayah, 2018; Stovall et al., 2019; Beese et al., 2022; Ma et al., 2023).

The appeal of individual-based methods applied to large-scale remote sensing data sets is not just in their ability to inventory huge numbers of trees. By capturing the 3D architecture of individual tree crowns, ALS data also provide a way to characterise tree growth along multiple axes, including height growth and lateral crown expansion (Lines et al., 2022). Vertical and horizontal crown growth are rarely recorded in field data, as they are much more complex and time-consuming to measure than stem diameters. However, they are arguably much more ecologically meaningful when it comes to capturing whole-plant growth strategies and how these vary with tree size, such as hypothesised shifts in biomass allocation away from height growth and towards crown expansion as trees approach maturity (Antin et al., 2016; Marziliano et al., 2019; Jucker et al., 2022; Laurans et al., 2024). Similarly, they are much more informative when it comes to understanding competitive interactions for light and space among neighbouring trees (Jucker et al., 2015; Taubert et al., 2015), as well as providing a way to quantify crown damage and dieback, which are strong predictors of tree mortality (Needham et al., 2022a; Zuleta et al., 2022). Another appeal of ALS data is that they contextualise the biotic and abiotic environment within which trees grow, such as their local competitive neighbourhood and topographic position within the landscape (Colgan et al., 2012; Swetnam et al., 2017; Beese et al., 2022; Ma et al., 2023). This provides an opportunity to not only quantify how demographic rates vary across the landscape but also ascribe this variation to underlying ecological drivers. Finally, a key selling point of individual-based methods is that they are directly comparable to how we monitor forests on the ground and how we represent them in forest dynamics models. This provides an opportunity to bridge the gap between field and remote sensing forest monitoring programmes and can help reduce major sources of uncertainty in forest dynamics models, such as those associated with tree mortality (Hubau et al., 2020; McDowell et al., 2020; Pugh et al., 2020).

Here, we use data from two ALS surveys acquired 9 yr apart in Australia's Great Western Woodlands (GWW) to capture the height growth, crown expansion, crown dieback and mortality of individual trees across 2500 ha of old-growth woodland habitat. These semi-arid woodlands are an ideal testbed for using repeat ALS data to quantify tree demographic rates at scale, as they are dominated by a small number of eucalypt species that form sparsely populated stands of single-stemmed trees. By developing a new pipeline for segmenting and matching tree crowns across ALS surveys, we were able to confidently identify and track the dynamics of 42 213 canopy-dominant trees across this landscape. Using these data, we set out to:

1. Determine how growth and mortality rates vary with tree size across the whole population. This allowed us to quantify which cohorts contribute most to biomass gains and losses, as well as characterise how trees adjust their crown growth strategies as they increase in size.
2. Model how tree growth and mortality rates vary across the landscape in relation to fine-scale topography and local neighbourhood competitive environment, so that we may better understand how demographic processes give rise to vegetation spatial patterns in dry forests.
3. Scale up tree-level demographic rates to community-level dynamics in aboveground biomass and canopy 3D structure. In doing so, we explored whether temporal changes in canopy 3D structure are predominantly driven by tree growth or mortality, and aimed to determine whether these old-growth woodlands currently operate as a net carbon sink or source in the absence of large disturbances from wildfires.