Résumé
Southern African drylands are not suited for the application of climate reconstruction methods based on surface samples. Methods based on the co-existence approach, while still in their early days, are really promising, particularly those using probability density functions (pdfs) that have proven particularly valuable in certain cases as they can be applied to a wide range of plants assemblages. Most commonly applied to fossil pollen data, their performance can be limited by the taxonomic resolution of the pollen data, as many species may belong to a given pollen-type. Consequently, climate information associated with a taxon cannot always be precisely identified, resulting in less accurate reconstructions. This can become particularly problematic in regions of high biodiversity, such as southern African botanical hotspots. The first part of this PhD thesis presents the development of a novel pdf-based climate reconstruction method adapted to the southern African context. The method, which comes along with a dedicated software pack- age entitled CREST, sorts out this diversity issue by taking into account the different climatic requirements of each species constituting the broader pollen-type: pdfs are fitted in two successive steps, with parametric univariate pdfs fitted first for each species (pdfsp) followed by a combination of those individual species pdfs into a broader single pdf to represent the pollen-type as a unit (pdfpol). The curve resulting from the multiplication of the pdfpol describes the likelihood of different climate parameters based on the co-existence of a given set of taxa, each being weighted according to its normalized pollen percentage. Three majors properties were derived from this continental-scale statistical analysis: 1) the method saturates when the number of species composing a pollen type becomes larger than 30-40 species, 2) the per- formance decreases with distance to the core of the climatic space and 3) climate variables that have a direct impact of plant life cycles are better reconstructed.We revisited 13 pollen sequences (selection based on their length, continuity, chronology and pollen diversity) from southern African literature with the CREST method. To offset the limited individual potential of those sequences, we developed a Monte-Carlo framework to create interpolated curves integrating uncertainties associated with the reconstructions and age-depth models and then stack those curves together to extract regionally consistent patterns. This reanalysis allows for the quantified reconstruction of a range of distinct climatic variables from this critical region, and provides significant insight into the nature of long-term climate change. Temperature reconstructions show strong coherency among all sites considered, and parallel southwest Indian Ocean SSTs. Reconstructions of the amount of summer precipitation since mid-MIS 3 (Marine Isotope Stage 3) 45,000 years ago indicate a dichotomy in the precipitation pattern between interior and northeastern South African sites. At the glacial-interglacial timescale, precipitation in northeastern sites shows strong similarities with the Indian Ocean SST records as well as with records from the large East African lakes. Entering the Holocene, precessional forcing becomes more important and a north/south rainfall dipole appears, with a demarcation line located somewhere between Lakes Tanganyika and Malawi (3-9°S). Sites from the interior, while clearly following a similar dynamic, also appear to be sensitive to additional factors, including the position of the southern Westerlies, which may interact with tropical systems to create tropical-temperate troughs. Our results shed light on the complexity of the mechanisms driving South African rainfall, and clarify several key elements of the current debate, including limitations of models relying on direct insolation forcing to explain long-term climate dynamics.