Final Report Summary - MIOVAT (Miocene Vegetation of the African Tropics (Project MioVAT))
Project MioVAT aims to investigate the long-term evolutionary history of the tropical forest and savannah biomes in West Africa. Laboratory work has been undertaken to release fossil pollen and spores from sediment samples using acid digestion and density separation (heavy liquid; sodium polytungstate) techniques. Fossil pollen and spores from the tropics of West Africa have been imaged using a variety of microscopy techniques. An image library of fossil pollen and spores derived from the Paleogene–Neogene rainforests of southern Nigeria has been compiled using transmitted light microscopy with brightfield illumination. Quercus (oak) and Picea (spruce) pollen has been imaged using transmitted light microscopy with brightfield illumination, confocal microscopy and scanning electron microscopy for the purpose of comparative microscopy work. Modern Poaceae (grass) pollen from herbarium sheets and fossil grass pollen from sediment samples has been imaged using transmitted light microscopy with brightfield illumination and SEM in order to gather data on pollen morphology across the grass family. Experiments have been undertaken comparing human and computational classifications of grass pollen. Computational methods for quantifying the complexity of grass pollen surface ornamentation have been extended to examine self-organized vegetation patterns in dryland ecosystems.
Project MioVAT has led to a new method of classifying pollen grains that is based on a combination of high-resolution imaging and computational image analyses. Experiments have been undertaken comparing these computational classifications with classifications produced human analysts. Computational methods achieve between 77.5% and 85.8% classification accuracy. Human analysts examining the same specimens achieve coverage of between 87.5% and 100% and identification accuracy of between 46.67% and 87.5%. The identification consistency of each human analyst ranged from 32.5% to 87.5%, and the proportion of duplicate image pairs that analysts missed ranged from 6.25% to 58.33%. Comparative microscopy work using Quercus, Picea and Poaceae pollen has shown that the taxonomic resolution of the pollen and spore fossil record can be increased considerably by using a combination of microscopy techniques that aim to recover morphological information from below the diffraction limit of light and computational image analyses. This in turn increases the range and depth of hypotheses that can be tested using the fossil pollen and spore record. These core methodological results that have already been published have provided a new computational approach to the description of 2-dimensional biological shapes, which has utility beyond the confines of the discipline of palynology. The results of this project have also highlighted and explored an emerging tension between the classification of biological objects by such computational methods, and classifications produced in a more traditional manner by human analysts.
Project MioVAT has led to a new method of classifying pollen grains that is based on a combination of high-resolution imaging and computational image analyses. Experiments have been undertaken comparing these computational classifications with classifications produced human analysts. Computational methods achieve between 77.5% and 85.8% classification accuracy. Human analysts examining the same specimens achieve coverage of between 87.5% and 100% and identification accuracy of between 46.67% and 87.5%. The identification consistency of each human analyst ranged from 32.5% to 87.5%, and the proportion of duplicate image pairs that analysts missed ranged from 6.25% to 58.33%. Comparative microscopy work using Quercus, Picea and Poaceae pollen has shown that the taxonomic resolution of the pollen and spore fossil record can be increased considerably by using a combination of microscopy techniques that aim to recover morphological information from below the diffraction limit of light and computational image analyses. This in turn increases the range and depth of hypotheses that can be tested using the fossil pollen and spore record. These core methodological results that have already been published have provided a new computational approach to the description of 2-dimensional biological shapes, which has utility beyond the confines of the discipline of palynology. The results of this project have also highlighted and explored an emerging tension between the classification of biological objects by such computational methods, and classifications produced in a more traditional manner by human analysts.