The Amazon's exceptional biodiversity has long been a subject of ecological and evolutionary debate. Haffer's refugia hypothesis, proposing that glacial periods fractured the rainforest into isolated refugia surrounded by savanna, has been highly influential but faces increasing criticism. Alternative hypotheses emphasize the role of rivers as isolating factors. Early climate modeling suggested a stable rainforest during the LGM, but these models lacked crucial interactions between CO2 levels, climate, and fire. Palynological evidence, while supporting a more stable Amazon, is limited by sparse sampling sites. Sato et al. (2021) used a more comprehensive approach, combining palynological data with DVM simulations, suggesting widespread savannafication driven by drier glacial conditions, low CO2 levels, and fire. However, their work didn't fully integrate model output and empirical data, leaving room for improvement in accuracy and interpretation. This study aims to refine past vegetation reconstructions by directly incorporating palynological data into DVM simulations via bias-correction, investigating the pattern and extent of forest and savanna fragmentation during the LGM and its implications for the refugia hypothesis. The research further explores the role of intermediate biomes, like woodland/tall savanna, in maintaining connectivity and facilitating species dispersal and diversification.

The refugia hypothesis, initially proposed by Haffer (1969), posits that during glacial periods, cooler and drier conditions fragmented the continuous Amazon rainforest into isolated pockets (refugia) separated by expanding savannas. This fragmentation, it is argued, restricted gene flow, leading to speciation and the high biodiversity observed today. This hypothesis, while influential, has faced significant challenges. Studies using palynological data (Bush et al., 2004; Mayle et al., 2009) and phylogenetic analyses (Bonaccorso et al., 2006) have presented contrasting views, suggesting either a less extensive fragmentation or even a largely stable rainforest. Alternative hypotheses have emerged, proposing the role of Amazonian rivers as significant barriers to gene flow (Pirani et al., 2019; Naka & Brumfield, 2018). Sato et al. (2021) attempted to reconcile these conflicting perspectives using a more comprehensive approach, utilizing a DVM model driven by paleoclimate data and integrated with palynological information, revealing significant savanna expansion but still with a considerable degree of forest connectivity. This study builds upon the work of Sato et al., refining the model and addressing some of its limitations.

This study utilizes the Land surface Processes and eXchanges (LPX) dynamic global vegetation model, driven by paleoclimate simulations from four global climate models (GCMs) for the LGM, as previously detailed in Sato et al. (2021). The study focuses on the ensemble reconstruction from Sato et al. (2021) that best matched available pollen data. Three key analytical approaches were employed: 1. **Biomisation:** A biomisation scheme, adapted from Prentice et al. and Ciais et al., was used to translate raw model bioclimatic outputs (growing degree days, fractional projected cover, vegetation height, evergreen/deciduous fraction, and tropical/temperate fraction) into biome classifications (tropical humid forest, tropical dry forest, etc.). A novel "woodland/tall savanna" category was introduced to capture intermediate vegetation types. 2. **Bias Correction:** A critical advancement was the incorporation of empirical palynological data from Marchant et al. (2009) into the DVM simulations through bias correction. This method improved model accuracy by adjusting model outputs to match pollen-based biome observations across multiple variables. Thin-plate spline surfaces were fitted to interpolate bias corrections between pollen core locations. 3. **Clustering:** A k-means clustering technique was employed on the bias-corrected fractional projected cover and height maps to identify regions of similar bioclimate. This objective clustering approach provides further insights into bioclimatic boundaries and connectivity patterns, complementing the expert-based biome classification. The study analyzed the results using three interpretations: pre-bias corrected, bias-corrected, and clustered biomes. Fragmentation was assessed by converting biome maps into polygons, counting the number of fragments and calculating the percentage of biome area outside the largest fragment. The study also includes sensitivity analyses to evaluate the robustness of the results to changes in biome thresholds. Finally, AI-generated imagery was created to illustrate the potential vegetation assemblages in each bioclimatic cluster.

The bias-corrected results reveal a substantially more fragmented Amazonian rainforest compared to the uncorrected simulations. The study identified at least five major forest fragments, with the largest located in the western Amazon basin. The bias-corrected simulations showed extensive savanna expansion, particularly in the northeast. However, critically, the results indicated that woodland/tall savanna vegetation connected many disparate forest fragments and savanna regions. This intermediate biome, with significant height and canopy cover, facilitated connectivity for generalist species. The bias-corrected simulations showed a significant increase in savanna area (5.51-7.06 million km² to 2.07-3.54 million km²) compared to uncorrected simulations. The clustering analysis identified four main bioclimatic regions: tall dense forests, dense forests/thickets, sparse/semi-arid vegetation, and short sparse desert/shrubland. The analysis revealed a large, interconnected cluster of dense vegetation across the Amazon, but also significant fragmentation in other vegetation types. Compared to the uncorrected data, the proportion of the forest area located outside the main fragment increased from 17.36% to 52.63% in the bias-corrected reconstructions. Sensitivity tests showed that changes in biome thresholds did not significantly alter the overall pattern of fragmentation, reinforcing the robustness of the findings.

The findings challenge the strict interpretation of the refugia hypothesis. While isolated forest patches exist in the bias-corrected reconstructions, these are connected by an extensive woodland/tall savanna biome. This intermediate vegetation type acted as a corridor, allowing gene flow and potentially explaining the lack of pronounced genetic breaks observed in some Amazonian species. The connectivity provided by this ecotonal biome may have facilitated range expansions during glacial periods for generalist species adapted to a broader range of conditions while specialist species may have been isolated within forest fragments, potentially leading to allopatric speciation. The study's refined reconstructions, integrating model outputs with empirical data, provide a more nuanced understanding of past vegetation dynamics. The existence of woodland/tall savanna suggests that past environments were not simply a dichotomy of closed forest and open savanna, but exhibited a more complex mosaic of habitats.

This study presents a refined understanding of Amazonian vegetation dynamics during the LGM, integrating empirical data with model simulations through a bias-correction approach. The results demonstrate a more fragmented forest landscape than previously suggested, yet also highlight the crucial role of an intermediate woodland/tall savanna biome in maintaining connectivity. This revised perspective on Amazonian paleovegetation significantly impacts our understanding of biodiversity evolution and challenges the simplistic interpretation of the refugia hypothesis. Future research could focus on further refining the model, incorporating additional proxy data, and exploring the ecological and evolutionary implications of the woodland/tall savanna biome in greater detail.

The study's reliance on a limited number of pollen core samples introduces uncertainty, particularly in areas distant from sampling sites. The model's inherent limitations in capturing the full complexity of past ecosystems may also influence the accuracy of reconstructions. Furthermore, the assumption of stationary climate in the model might not fully capture the dynamic nature of paleoclimates. It is also worth noting that the AI-generated imagery provides a plausible visualization of the vegetation types but may not perfectly represent the exact flora.