Identifying wetland type from pollen assemblages using machine learning

Wetlands act as carbon sinks worldwide; thus, their preservation and restoration help slow global warming. Wetland types (swamps, marshes, fens, bogs) differ in terms of their rates of carbon uptake, burial and release, and stability of soil organic matter. Understanding long-term wetland development through paleoenvironmental records improves predictions of the effects of climate change and human impact on wetland functions in different types of wetlands. In order to reconstruct past environments and understand their climate effects, pollen is used as an indicator of vegetation changes. However, wetland pollen is widely distributed and often identified at low taxonomic resolution (e.g., Poaceae or Grasses), where upland pollen taxa gain more focus in the paleoclimate reconstruction. Therefore, this research aims to present new methods to assign modern pollen assemblages from previous studies to specific wetland types using machine learning in order to classify each taxon into a wetland type. We develop learning models to establish a relationship between modern pollen counts and wetland types using machine learning techniques. A collection of marsh datasets ranging from high to middle to low marsh, and a modern pollen dataset from boreal peatlands with open fen, treed fen, open bog, and treed bog classifiers are used in our training. The two datasets are highly non-linear, small in number, and have imbalanced sample counts in each class. Achieving a high-performance estimator with a very small number of samples is challenging. Employing random forest and deep learning algorithms, we reach over 80% prediction accuracy. We also compare machine-based calculations and manually selected features in our model using feature reduction. The high-performance model is applied to fossil records and used to estimate wetland types in the past. Classification of these taxa into a wetland type using modern pollen datasets databases, usually with a significant number of taxa (features) from the modern pollen records, is challenging and time-consuming. Thus, machine learning can be a reliable tool for classifying wetland types in paleoenvironmental studies.