Discussion
Our integrated modelling approach helps fill a gap in ecosystem spatial
modelling capacity. We extend Ferrier and Guisan’s (2006) well
established ‘assemble and predict together ’ strategy for
community-level distribution modelling to predict ecosystem spatial
patterns. Here, we model joint biotic and abiotic responses as a
function of environmental gradients. Thereafter, we apply spatial
turnover among different combinations of modelled biotic and abiotic
responses as a basis for identifying ecosystem patterns. This effort
marks one of the first spatial ecosystem models to integrate of biotic
and abiotic responses. The novel approach helps illuminate relationships
among biotic and abiotic ecosystem properties, and their drivers. It
also provides a basis for resolving the collective contributions biotic
and abiotic model responses make to ecosystem spatial heterogeneity
across landscapes. Our approach can be applied in other regions to
provide conservation planners and decision makers with a tool to predict
how and where ecosystems vary, and a means to help understand the
origins of these patterns. The approach is well suited to regions where
the density of survey sites is low relative to the grain of spatial
turnover in the targeted ecological entity (e.g., community, ecosystem)
(Ferrier et al. 2007). It is also ideal when the objective is to predict
continuous spatial variation in ecosystem makeup and to apply those
predictions for mapping emergent biodiversity patterns (Basquill and
Leroux 2023).
The ecosystem is central to recent global conservation agreements, such
the Kunming-Montreal Global Biodiversity Framework, and new scientific
guidance is available to implement these and other ecosystem targets
across, global, national and sub-national extents (Nicholson et al.
2024, Venegas-Li et al. 2024). A critical impediment is that few
sub-global regions have adequate ecosystem mapping (Xiao et al. 2024),
and statistical models required to establish those maps are rare (Geary
et al. 2020). To address these challenges, previous terrestrial
approaches have included modelling the distribution of pre-classified
ecosystem types (e.g., Simensen et al. 2020, Naas et al. 2024), or
proxies including pre-classified vegetation community types (e.g., Comer
et al. 2020, Jiménez‐Alfaro et al. 2023) or remotely sensed land cover
classes (e.g., Murray et al. 2022). Most examples have been conducted at
coarser spatial grains and correspond with Ferrier and Guisan’s (2006)
‘assemble first, predict later ’ strategy for modelling
biodiversity patterns. With this strategy, individual observations –
species for example – are assembled (e.g., through numerical
classification) into discrete units, such as community types, and those
units are modelled in space. Following the ‘assemble and predict
together ’ strategy, our alternative approach fills a gap for modelling
ecosystem patterns at the landscape extent. In addition, we predict at a
fine spatial grain. This grain matches the spatial resolution where
ecosystem restoration, and much ecosystem-based management and
conservation, is implemented (e.g., Mäkelä et al. 2012, Pressey et al.
2013, Aubin et al. 2024).
Different analytical approaches have distinct strengths, and their
characteristic applications reflect varied modelling data and intents
(Deschamps et al. 2023, Naas et al. 2024). Our adaptation of Ferrier and
Guisan’s (2006) ‘assemble and predict together ’ strategy enabled
us to parse the shared and independent components and determinants of
ecosystem heterogeneity. More specifically, it allowed us to predict
biotic, abiotic, and ecosystem responses to diverse environmental
circumstances, which we operationalized in three separate models. Two of
our models represent lower levels of ecological organization, which
Levins (2011) contends have a controlling influence on ecosystem pattern
and process. Predictions from lower ecological levels also allows
ecosystem conservation practitioners to represent biotic and abiotic
properties in model-informed conservation plans. Here, these
distinctions may be necessary for managing or restoring specific groups
of ecosystem components or services (e.g., wildlife habitat - Van der
Biest et al. 2020; carbon sequestration and storage - Ameray et al.
2021; soil properties - Rader et al. 2022). Our modelling framework
provides this functionality, enabling researchers to predict whole
ecosystems and their parts, while identifying potential mechanisms.
Our foremost predictors of ecosystem dissimilarity were all
vegetation-based, namely leaf area index, softwood basal area, and
normalized difference vegetation index (Table 1). Vegetation predictors
(e.g., canopy height, normalized vegetation difference index) similarly
explained the largest proportion of variation in forest ecosystem
distribution models of Norway (Naas et al. 2024). In distribution models
of vegetation-based ecosystem proxies (e.g., Ponomarenko et al. 2019,
Lee et al. 2021) spectral vegetation indices (e.g., normalized
vegetation difference index) and remotely sensed vegetation cover are
frequently the primary predictors. These mutual findings suggest the
potential effectiveness of employing vegetation canopy features – which
represent a fraction of total ecosystem composition and structure – to
predict forest ecosystems.
The most influential predictors in our biotic and ecosystem models were
identical. This might imply the predominant drivers of biotic properties
mirror those shaping whole ecosystems – including constituent facets of
abiotic diversity. An alternative interpretation is that the relative
influence of abiotic predictors on ecosystem variation has been masked
in our ecosystem formulation. Results of our independent biotic and
abiotic response models lend relevant evidence. In the abiotic response
model, terrain ruggedness and depth to water, both frequent topographic
predictors in biodiversity and geodiversity models (Dilts et al 2023,
Toivanen et al. 2024), were the top two predictors, accounting for 80%
of total deviance explained (Figure 4). Neither of these predictors were
important in the biotic or ecosystem response models. One explanation
for these disparate findings may lie in our abiotic response model’s
performance, which was the lowest of the three builds. This model’s
lower relative performance may indicate new predictors of abiotic
variation are warranted. Furthermore, including these same new
predictors in an ecosystem response model could shift the ratio of
meaningful predictors more evenly among biotic and abiotic predictors.
Indeed, commonalities between the biotic and ecosystem response models,
and the intermediate performance of the latter model, suggest these two
formulations are more strongly favouring biotic properties and their
respective predictors.
Generally, our models are exploratory and could benefit from empirical
testing with out of sample data from different study areas or time
frames (Tredennick et al. 2021). The absence of these data highlights a
limitation of our adaptive approach (sensu Holling and Allen 2002). This
approach is common in models where relationships between predictors and
response variables are poorly understood (Tredennick et al. 2021,
Planque et al. 2022), such as they are between ecosystem properties and
their determinants (Soranno et al. 2019). Empirical testing could also
lead to hypothesis generation and theory development, particularly for
mechanisms of ecosystem assembly, which are inadequately substantiated
(Levin 1998). Model outcomes could also be employed for making secondary
predictions to explicitly link pattern and process (Gallagher et al.
2021). The framework we develop lends itself well to these objectives.
It can be rapidly fit to large datasets, allowing for the development
and testing of nested models for successive resolution of ecosystem
patterns and drivers at various scales and levels of complexity.
One strategy for strengthening our ecosystem build is to draw more
explicitly from geodiversity modelling and its potential for improving
biodiversity predictions. Recognition of the interplay between
biodiversity and geodiversity – the diversity of Earth’s abiotic
features and processes, including climate – has risen sharply, partly
in response to rapidly changing global conditions (Schrodt et al. 2024).
Efforts to bridge these two disciplinary foci have emerged from both
biodiversity (Vernham et al. 2023) and geodiversity (Alahuhta et al.
2024, Toivanen et al. 2024) research streams. The ecosystem is an
apparent focal unit for synthesizing across disciplines, for
geodiversity is integral to ecosystems (Richter and Billings 2015,
Holling 1992, Ochoa-Hueso et al. 2021). Yet, few ecosystem models
jointly incorporate facets of biotic and abiotic diversity as responses
(Basquill and Leroux 2023). While we strived to adequately represent
abiotic responses in our ecosystem model, the inherent coarse spatial
grain of many geodiversity predictors – including geology and climate
(e.g., Hjort and Luoto 2012, Read et al. 2020) – precluded their use in
our study. We selected predictors to match our survey grain following
best practices for biodiversity distribution modelling (Chauvier et al.
2022). Although omitting coarse-grain predictors in our models may seem
like a modelling impediment, it raises an opportunity for exploring
ecosystem scaling relationships.
Ecosystems occur at all spatial scales (Fritsch et al. 2020) and scales
selected for modelling correspond to study objectives (Geary et al.
2020). The fine resolution (10 m) in our models reflects the spatial
grain where many topographically controlled terrestrial ecosystem
properties (e.g., vegetation – Moeslund et al. 2013; soil – Seibert et
al. 2007; water – Detty and McGuire 2010; organic matter – Burton et
al. 2011) are strongly expressed. This grain coincides with the
fine-grained mapping needed for much natural resource management
(Pressey et al. 2013, D’Urban Jackson et al. 2020, Senf 2022) and
applied forecasting (Dobrowski 2011). However, ecosystem properties may
also be driven by processes occurring at other scales (Holling 1992).
For example, climatic processes affect biodiversity from macro to
microscales (Coelho et al. 2023, Kemppinen et al. 2024). The paucity of
microscale climate data has limited their availability for fine-grained
studies (Kling et al. 2024) and resulted in possibly misleading
biodiversity predictions (Slavich et al. 2014, Manzoor et al. 2018).
Notwithstanding this data deficit, one solution for investigating
scaling relationships is to systematically examine the influence of
spatial grain on model outcomes (e.g., Guisan et al. 2007). Numerous
methods are available for altering grain to advance understanding of
scaled relationships across landscapes (Graham et al. 2019, Markham et
al. 2023). Our modelling framework would lend itself well to this
inquiry, providing a means to test how scaling is affected by grain and
level of ecological organization (i.e., biotic, abiotic, and ecosystem
levels). Another benefit of our framework is that it could be coupled
with separate spatial models to predict assemblages of ecosystem
properties.
Our GDM-based framework is a strategy for predicting shifting
combinations or turnover of ecosystem properties, analogous to GDM
models of community turnover. Working from this premise, we posit the
comparable but distinct value of predicting spatially structured
assemblages of ecosystem properties. This objective could also be
operationalized by extending community-level modelling techniques, such
as joint species distribution modelling (jSDM; Basquill and Leroux
(2023). jSDM is a community modelling technique (Pollock et al. 2014)
which can be applied to predict biotic composition (Franklin 2023). The
two adaptations of community-level modelling techniques – GDM in the
present study and jSDM – offer complementary approaches for predicting
ecosystem patterns. Here, we suggest they could be coupled to predict
spatially structured 1) assemblages of ecosystem properties (via jSDM)
and 2) continuous shifts in those assemblages (via GDM). Similar
pairings of GDM and jSDM were employed for predicting taxonomic and
functional composition and turnover of European peat bogs (Robroek et
al. 2017), and the assembly and turnover of meiofaunal communities
across Denmark (Macher et al. (2024).