Machine Learning Models for Swedish Waterbird Populations in Areas with Sparse Sampling

Sammanfattning

Every year, the Swedish Bird Survey at Lund University performs a census of waterbirds throughout Sweden. These records are compiled in a yearly report by the Swedish Bird Survey, with the purpose of monitoring and detecting trends in the populations of Swedish birds. The motivation and goal of this project is to explore if machine learning methods can support these conservation efforts. Firstly by creating spatial models that can supplement missing bird counts in areas with sparser sampling, and hence assist in the bird survey. Secondly by creating temporal models to predict future populations. We also aim to analyze the importance of the model features, and search for predictors explaining the trends. For both the spatial and temporal task, we used a Poisson regression model as a baseline. We then explored extensions of this model, such as generalized additive models, in addition to treebased machine learning methods, such as LightGBM. We subsequently compared the model performances and found that LightGBM was the overall best performing model for both tasks.