Paleobotanical data enhances past predicted distribution of an evergreen woody Genus (Arbutus L.)

Hindcasting is a widely used method to predict the past distribution of plant species. However, the general lack of past observations hampers the accuracy of the predicted suitable areas. Here we provide the modelled distribution for the genus Arbutus in the Macaronesian-Mediterranean region, including confirmed paleobotanical occurrences to calibrate the ecological model for seven bioclimatic periods from the Last Glacial Maximum to the late Holocene and to compare them with the traditional hindcast method. The ensemble distribution model was based on high-resolution occurrence records and bioclimatic variables respectively retrieved from literature and databases. The Spearman indicator enabled the selection of variables with lower correlation. Past data consist of calibrated pollen and macrofossil records. First, we trained the model for current conditions and then generated hindcasts for each past period. We evaluated the accuracy of hindcasts for each period using a confusion matrix (modelled vs observed). Secondly, we trained models with fossil occurrences for each past period and then validated them through the same matrix. We also assess the Spatial Coherence among different methods over time by applying Jaccard’s similarity index. Results show differences in predicting the suitable area among the different methods. The Balanced Accuracy shows a high magnitude discrepancy between Hindcasted projections and projections calibrated for each period (Presence-Only; Presence-Absence; Presence-Absence+ pseudoabsence), and accordingly, the Spatial coherence presents a chronologically increasing reliability. This work highlights the importance of including paleobotanical records to increase the accuracy of species' past projections and improve the quality of historical biogeography questions.