One climate pressure can create different outcomes
Climate change is often discussed as a single global threat to wildlife, but animals do not experience environmental change in the same way everywhere. A rise in temperature may make one location less suitable for a species while improving conditions somewhere else.
A new study published in Nature Communications on 16 October 2026 used explainable artificial intelligence to examine that geographical complexity for 1,992 terrestrial mammal species worldwide.
The researchers found that climate extremes can create intense but geographically restricted effects, while temperature-related factors are especially important in explaining projected changes in habitat suitability.
Crucially, the same environmental driver could have opposite associations in different regions. The results challenge conservation approaches that assume one climate variable has a uniform effect across a species’ entire range.
The findings concern projected habitat suitability, not a count of animals that will definitely survive or disappear. They are designed to improve understanding of risk and opportunity, rather than predict the fate of individual populations with certainty.
Why global averages can conceal local risks
Global temperature trends are essential for understanding climate change, but an average cannot describe every ecological consequence.
Different populations occupy landscapes with different elevations, rainfall patterns, vegetation and seasonal conditions. Their exposure to environmental pressures therefore varies.
A species living near the warm edge of its range may experience additional warming differently from populations closer to its cooler range boundary.
Changes in rainfall can also interact with habitat structure and land cover, creating different conditions even in locations with similar annual temperatures.
Traditional species-distribution models can estimate where environmental conditions are suitable, but their predictions are often presented as aggregate maps.
Such maps may reveal where suitability increases or decreases without explaining which environmental variables contribute most strongly to the change.
The new study attempts to make those predictions more interpretable by examining the contributions of individual drivers at particular locations.
How the researchers analysed nearly 2,000 mammals
Lei Song, Amy E. Frazier, Peter Kedron and colleagues examined projected changes in the distributions of 1,992 terrestrial mammal species.
The team combined species-distribution modelling with an explainable machine-learning technique known as Shapley Additive Explanations, or SHAP.
The analysis considered 16 environmental drivers covering average climate conditions, climate variability, climate extremes and land cover.
Rather than treating the modelling system as a black box, the researchers examined how individual variables contributed to predicted suitability.
The approach allowed them to compare contributions across locations and across different projected periods.
It also helped distinguish between a variable that consistently creates risk and one whose influence changes direction from place to place.
These are model-based associations. The authors explicitly state that the framework was not designed to establish causation.
What explainable AI adds
Many machine-learning models can identify complex patterns but provide limited explanations for their predictions.
SHAP methods help address that problem by estimating the contribution of each input variable to a model’s output.
In this research, the method was applied to species-distribution models to investigate why predicted habitat suitability changes.
For example, a model might indicate that an increase in a temperature-related variable makes a particular area less suitable for one species.
In another region, the same variable could contribute positively to the predicted suitability for that species.
Examining these local contributions helps researchers move beyond a single global ranking of environmental threats.
However, a SHAP contribution explains a prediction made by a model. It does not prove that the environmental variable directly caused an observed change in wildlife populations.
That distinction matters when translating machine-learning results into conservation decisions.
Climate extremes produced concentrated effects
One major result was that climate extremes generated strong but spatially restricted impacts in the modelled distributions.
An extreme event can place intense pressure on wildlife in a particular location without affecting all parts of a species’ range equally.
This differs from average climatic conditions, which may shape suitability across much broader areas.
The finding suggests that maps showing only long-term average warming can overlook important localised pressures.
Extreme heat, unusually dry conditions or other departures from typical climate can create short-lived or recurring challenges.
The study assessed environmental drivers in a modelling framework rather than directly observing the survival of mammals during particular extreme events.
It therefore identifies potential geographical patterns of exposure and suitability, not confirmed mortality caused by individual weather episodes.
Nevertheless, the distinction between broad and concentrated pressures is important for planning where monitoring and conservation attention may be needed.
Temperature-related drivers stood out
Temperature-related environmental factors emerged as dominant contributors to the projected exposure patterns across the mammals examined.
The researchers also reported that specialist species could face particularly important effects from these drivers.
Specialists are animals with relatively narrow ecological requirements, such as dependence on particular resources or environmental conditions.
Species with restricted tolerances may have fewer suitable alternatives when conditions shift.
However, specialists are not a single uniform group. Their responses can vary according to location, physiology, habitat availability and other pressures.
The modelled prominence of temperature does not mean precipitation, land cover or climate variability are unimportant.
Rather, it indicates that temperature-related factors were particularly influential within this study’s global modelling framework.
Conservation assessments still need to consider how several drivers interact at local and regional scales.
The same driver can signal risk and opportunity
Perhaps the most useful conceptual finding is that an environmental driver need not have the same effect everywhere.
The researchers identified cases in which the direction of a driver’s modelled contribution changed across space or time.
An environmental change associated with reduced suitability in one area could be associated with increased suitability elsewhere.
That does not mean the positive and negative effects cancel each other out for conservation purposes.
Newly suitable habitat may be inaccessible because of roads, farmland, fragmented ecosystems or geographical barriers.
Animals may also be unable to disperse quickly enough to occupy an emerging suitable area.
Likewise, a modelled improvement in climatic suitability does not guarantee food, shelter or successful reproduction.
Understanding the location and direction of change is therefore more useful than assuming a global net gain or loss tells the whole story.
Two ways of diagnosing environmental change
The study introduced two ecological diagnostics called directional turnovers and favorability shifts.
Directional turnover describes situations in which the contribution of an environmental driver switches between favourable and unfavourable, or vice versa.
Favorability shifts capture changes in how strongly a driver contributes even when the direction does not reverse.
Together, these measures offer a more detailed picture than a simple map of predicted habitat loss.
For instance, an environmental factor may remain unfavourable in a location while becoming substantially more harmful over time.
In another location, a driver may cross from a positive to a negative contribution, signalling a change in its relationship with suitability.
These diagnostics are analytical tools for interpreting model outputs, not direct measurements of animal behaviour.
Their value lies in helping researchers identify where environmental relationships appear to be changing most meaningfully.
Why conservation planning needs local evidence
Conservation authorities frequently face limited budgets and must decide which habitats or populations need attention first.
Global assessments can identify broad priorities, but they cannot substitute for detailed regional information.
The study suggests that driver-specific maps could help distinguish areas threatened mainly by temperature changes from areas where land cover or other conditions are more influential.
That information may support more targeted monitoring, restoration and habitat-protection strategies.
For example, a region with suitable future climate but severe fragmentation may need a different response from an area where climate suitability itself is deteriorating.
However, the paper does not experimentally test particular conservation interventions or calculate their costs.
Decision-makers would still need information about species abundance, dispersal, habitat condition and the feasibility of local management.
Explainable AI can inform those discussions, but it should not be treated as a replacement for ecological fieldwork.
What the findings could mean for African wildlife
Africa contains diverse mammal communities occupying savannas, forests, mountains and arid landscapes.
These ecosystems experience different combinations of temperature, rainfall and land-use pressures.
The global study includes terrestrial mammals from around the world, but its findings should not be interpreted as a site-specific forecast for any particular South African reserve.
Local management decisions require information about the species present, environmental trends and the condition of surrounding habitats.
The finding that one driver can produce opposite suitability effects across locations is especially relevant in landscapes with strong ecological gradients.
Wildlife conservation must also consider whether animals can move between habitats as conditions change.
Protected areas may face challenges if suitable conditions shift beyond their boundaries or if surrounding land use limits movement.
These are practical questions that the modelling framework helps highlight, even though it does not resolve them by itself.
The limitations of a modelling study
The analysis relied on species-distribution models and projected environmental conditions rather than controlled experiments.
Predictions depend on the available occurrence data, environmental variables and modelling assumptions.
Some species are monitored more thoroughly than others, which can affect the quality of the information used.
Habitat suitability is not identical to population size, survival, reproductive success or extinction probability.
Species may also adapt behaviourally, shift their activity patterns or respond to interactions with other species in ways not fully represented by the models.
SHAP-based explanations reveal which inputs shape model predictions, but correlated environmental variables can complicate interpretation.
The authors explicitly caution that their framework is not intended to establish causal relationships.
These limitations are reasons to combine modelling with field observations and independent ecological evidence.
Why this research matters
Climate change is reshaping ecosystems, but conservation decisions depend on understanding more than global averages.
By analysing 1,992 terrestrial mammal species and 16 environmental drivers, the researchers showed how explainable machine learning can uncover geographical differences in projected ecological pressures.
Temperature-related factors were prominent, while climate extremes produced concentrated effects that broad averages can miss.
The same driver sometimes contributed positively in one location and negatively in another, highlighting the complexity of future habitat suitability.
The new diagnostics provide a way to identify changes in both the direction and strength of environmental influences.
The central message is not that climate change will benefit mammals overall. It is that ecological risk is spatially uneven and requires regionally informed responses.
More detailed forecasts, supported by local observations, may help conservation planners recognise emerging threats and opportunities before they are obscured by global summaries.
Source Information
Study: Explainable artificial intelligence reveals spatially divergent effects of global change on mammals.
Authors: Lei Song, Amy E. Frazier, Peter Kedron, Diogo S. A. Araujo, Diyang Cui and colleagues.
Journal: Nature Communications.
Published: 16 October 2026.
Methods: Explainable machine-learning analysis of species-distribution models for 1,992 terrestrial mammal species, evaluating 16 environmental drivers with Shapley Additive Explanations.
