Modelling reveals the effect of climate and land use change on Madagascar's chameleons fauna

The study addresses how climate change and land use/land cover (LULC) conversion jointly shape the future distribution and survival prospects of Madagascar’s chameleons, a predominantly endemic and often narrowly distributed group. Madagascar has experienced decades of deforestation and ongoing land conversion, while also facing climate change impacts that may variably contract or expand suitable habitats. Predicting outcomes for chameleons is challenging due to high endemism, rarity, and limited occurrence data for many species, which complicates species distribution modeling (SDM). The authors aim to: (i) extend ENphylo, a phylogeny-informed approach, to model even extremely rare species (down to two occurrences), (ii) project present and future distributions for Madagascar’s chameleons under mild and severe climate and LULC scenarios, and (iii) quantify and map the additive, antagonistic, synergistic, and single-driver (“only climate” or “only LULC”) effects of climate and land conversion on species loss and gain across Madagascar.

Prior work highlights Madagascar as a biodiversity hotspot under severe pressure from deforestation and land use change, with protected areas expanding yet not fully buffering ongoing habitat loss. Traditional SDMs typically require ≥10 occurrences, leading to exclusion of many rare species from projections; earlier studies on Madagascar’s plants and reptiles adopted such thresholds, with chameleons sometimes modeled only with ≥20 points. Ensembles of small models (ESMs) and other methods have been proposed for rarity, but performance and minimum data needs remain limiting. ENphylo was introduced to leverage phylogenetic signal in climatic preferences to impute niches for poorly sampled species, demonstrating strong performance with 10–20 points. However, extremely rare taxa (≤5 points) still posed challenges. Previous projections for Madagascar suggest that while climate change can cause both contractions and expansions, land conversion is often a stronger, more uniformly negative driver. The literature also emphasizes uncertainties stemming from socioeconomic trajectories, habitat connectivity, and the limited dispersal abilities of many small vertebrates, including chameleons.

Occurrence data: Chamaeleonidae occurrences were downloaded from GBIF (with coordinates; filtered to material citation and machine/human observations; accuracy ≥0.01°; duplicates/unrealistic records removed; non-African/Mediterranean data excluded), yielding 17,170 occurrences for 151 species. After removing duplicate occurrences per cell, 6,915 remained. A composite phylogeny was assembled (Tonini et al. and Giles et al.) via RRphylo’s tree.merger; species lacking geographic data, single-cell occurrences, or conflicting phylogenetic info were excluded, resulting in 134 species (56 in Madagascar). Environmental predictors: CHELSA v2.1 bioclimatics (1 km; baseline 1981–2010; futures 2071–2100) for SSP1-2.6 (mild) and SSP5-8.5 (severe) across three CMIP6 GCMs (GFDL-ESM4, MRI-ESM2-0, IPSL-CM6A-LR). Land Use/Land Cover (LULC) predictors included seven categories, represented as Euclidean distance rasters. Variables were checked for multicollinearity (Pearson r > 0.7 threshold), retaining 11 predictors: BIO4, BIO5, BIO7, BIO13, BIO15, and distances to water bodies, forests, grasslands, barren areas, urban areas, and croplands. Resolution was 1 km over Madagascar. Modeling approaches: Three SDM strategies based on sample size: (i) ENphylo for species with <15 occurrences; (ii) ESMs for 15–30 occurrences (all pairwise combinations of predictors); (iii) traditional SDMs (MaxEnt, Random Forest, GLM) for >30 occurrences. For GLM, quadratic terms and interaction level = 1 were used; other algorithms used biomod2 defaults. ENphylo used 10,000 background points per species. ENphylo imputes marginality and specialization axes for rare species via phylogeny after fitting ENFA to well-sampled taxa; at least five presence-like points are needed to convert imputed niches to Mahalanobis distance-based suitability. To extend to <5 true occurrences, the authors generated pseudo-presences adjacent to true presences (knight-move cells), then selected the closest in climate/LULC space (via vector angle similarity), adding n pseudo-presences so that reference + pseudo-presences = 5. Phylogenetic uncertainty was addressed by testing 50 modified trees (swapOne in RRphylo), selecting the replicate with the best AUC. Validation and projections: For all approaches, 80/20 bootstrap cross-validation with replacement was repeated 10 times. AUC, TSS, and Boyce index were computed; models with AUC < 0.7 were removed. Model averaging for ESMs/SDMs weighted projections by AUC. Present and future projections were made under three scenario types: (i) dynamic climate with LULC held constant, (ii) dynamic LULC with climate held constant, and (iii) both dynamic (LULC-climate). Predictions were binarized using three thresholds: equalize sensitivity and specificity (SensSpec), maximize TSS (MaxSens+Spec), and minimum training presence (TenPerc). A dispersal constraint of 1 km per year (2010–2070: 60 km) was applied by buffering the minimum convex polygon of current occurrences before cropping current and future binary maps. Post-processing: Binary maps were stacked to compute species richness, loss, and gain per grid cell. Species loss and gain were compared between current and future scenarios to partition driver interactions into synergistic (c > a+b), additive (c ≈ a+b), antagonistic (various c < a/b conditions), only climate (a=0, b>0), and only LULC (a>0, b=0), where a = dynamic land use, b = dynamic climate, c = dynamic LULC-climate. Percentages of interaction types were summarized for mild and severe scenarios. Overall, 45 projections per species were produced by combining thresholds, GCMs, and mild/severe scenarios for climate and LULC.

Model performance and autocorrelation: Spatial autocorrelation in ensemble residuals was negligible (mean Moran’s I = −0.13, sd = 0.06; 14% significant replicates). Excluding species with <10 occurrences for performance reporting, ENphylo achieved mean AUC = 0.918 (sd = 0.106), TSS = 0.703 (sd = 0.166), Boyce = 0.394 (sd = 0.122). ESMs: AUC = 0.953 (sd = 0.033), TSS = 0.871 (sd = 0.086), Boyce = 0.827 (sd = 0.067). Traditional SDMs: AUC = 0.964 (sd = 0.033), TSS = 0.864 (sd = 0.100), Boyce = 0.965 (sd = 0.04). Species loss/gain and driver effects: Across Madagascar, future projections showed extensive areas of loss and gain, with magnitudes increasing from mild to severe scenarios. LULC change was the dominant driver at broad spatial scales, while climate change effects were more local. Table 1 (averages across thresholds and GCMs) showed: - Dynamic climate: average loss 0.163 (mild), 0.275 (severe); average gain 0.060 (mild), 0.067 (severe); max loss 6.000 (mild), 8.889 (severe); max gain 2.444 (mild), 2.000 (severe). - Dynamic land use: average loss 0.749 (mild), 0.737 (severe); average gain 0.722 (mild), 0.573 (severe); max loss 6.667 (mild), 7.667 (severe); max gain 8.000 (mild), 8.000 (severe). - Dynamic land-climate: average loss 0.823 (mild), 0.890 (severe); average gain 0.677 (mild), 0.470 (severe); max loss 8.889 (mild), 11.222 (severe); max gain 6.889 (mild), 6.111 (severe). Interaction types (Table 2, averaged across thresholds and GCMs): - Mild scenario: Loss: synergistic 1.70%, additive 2.70%, only climate 5.28%, antagonistic 15.13%, only LULC 75.19%. Gain: synergistic 3.62%, additive 0.86%, only climate 6.57%, antagonistic 3.68%, only LULC 85.27%. - Severe scenario: Loss: synergistic 1.59%, additive 4.04%, only climate 8.16%, antagonistic 19.16%, only LULC 67.06%. Gain: synergistic 6.21%, additive 0.92%, only climate 9.51%, antagonistic 4.62%, only LULC 78.74%. Spatial patterns and species at risk: Highest projected biodiversity loss is concentrated in the western and north-western dry deciduous forests; highest turnover (losses and gains) is in eastern lowland forests. Approximately 10–20 species are predicted to lose >90% of their current range under future scenarios, including several Brookesia (e.g., B. desperata, B. karchei, B. micra, B. tristis), Calumma (e.g., C. amber, C. guibei, C. ambreense, C. nasutum, C. fallax, C. peltierorum, C. boettgeri, C. globifer), and Furcifer (e.g., F. petteri), as well as B. decaryi, B. brunoi, and Furcifer willsii. Some widespread species (e.g., Furcifer oustaleti, F. rhinoceratus; Calumma parsonii, C. crypticum, C. brevicorne, C. oshaughnessyi; Brookesia superciliaris) may gain range. However, chameleons’ low dispersal (assumed 1 km/year) and uncertain habitat connectivity render potential gains largely hypothetical absent assisted movements. Overall, land conversion is predicted to override climate change effects in determining future suitability patterns.

The results directly address the central question of how climate and LULC changes, individually and in combination, will affect Madagascar’s chameleons. Land conversion emerges as the predominant driver of future habitat loss, with “only LULC” effects accounting for the majority of both losses and gains across scenarios, while synergistic effects are rare and antagonistic effects are more notable for losses. Spatially, severe losses are concentrated in the western and north-western dry deciduous forests, whereas eastern lowland forests are projected to experience high turnover. Although some widespread species may find additional suitable areas, most projected gains are uncertain because of limited dispersal and potentially poor habitat connectivity. These findings reinforce the notion that human land use pressures are more immediately threatening to chameleons than climate change alone and underscore the importance of land management, protected area effectiveness, and corridor planning for conservation. Extending ENphylo to extremely rare species enables inclusion of taxa that are often omitted from future projections, providing a more complete basis for conservation prioritization.

This study extends the ENphylo framework to model extremely rare species (down to two occurrences) by combining phylogenetic niche imputation with carefully selected pseudo-presences, enabling robust projections for 56 Madagascar chameleon species under multiple climate and land use scenarios. The analysis shows that land use change will largely dominate future habitat outcomes, driving substantial losses particularly in the island’s western and north-western dry deciduous forests. A notable fraction of species is projected to undergo severe (>90%) range contractions, placing them at high risk of extinction, while any potential range gains are uncertain due to limited dispersal and connectivity. The work provides actionable insights for conservation focusing on mitigating land conversion, strengthening and connecting protected areas, and managing corridors that could facilitate range shifts. Future research should refine dispersal and connectivity assumptions, integrate dynamic corridor availability, improve socio-economic scenario realism, and continue enhancing modeling approaches for data-poor species.

Key limitations include: (i) reliance on sparse occurrence data for many species, with validation metrics like AUC being sensitive to prevalence and small sample sizes; (ii) the use of pseudo-presences to reach a minimum of five presence-like points may bias niches toward narrowness and does not guarantee true occurrence; (iii) assumptions of a uniform dispersal rate (1 km/year) and effective habitat permeability/corridors may overestimate realized gains; (iv) uncertainties in future LULC and climate scenarios, as well as socio-economic trajectories and protected area management, can affect projections; (v) possible residual spatial autocorrelation and phylogenetic uncertainty, though addressed, cannot be fully eliminated; and (vi) limited ability to capture fine-scale microhabitat and biotic interactions which can influence chameleon distributions.