Explainable machine learning for modelling of net ecosystem exchange in boreal forest

There is a growing interest in applying machine learning methods to predict net ecosystem exchange (NEE) based on site information and climatic parameters. In case of successful performance, it could give an excellent opportunity for gapfilling or upscaling, i.e., extrapolation of results to times and sites for which direct measurements are unavailable. There exists already quite an extensive body of research covering different seasons, time scales, number of sites, input parameters (features), and models. We apply four machine learning models to predict NEE of boreal forest ecosystems based on climatic and site parameters. We use data sets from two stations in the Finnish boreal forest and model NEE during the peak growing season and the whole year. Using Explainable Artificial Intelligence methods, we compare the most important input parameters chosen by the models. In addition, we analyze the dependencies of NEE on input parameters against existing theoretical understanding on NEE drivers. We show that even though the statistical scores of some models can be very good, the results should be treated with caution especially when applied to upscaling. In the model setup with several interdependent parameters ubiquitous in atmospheric measurements, some models display strong opposite dependencies on these parameters. This behavior might have adverse consequences if models are applied to the data sets in future climate conditions. Our results highlight the importance of Explainable Artificial Intelligence methods for interpreting outcomes from machine learning models, in particular, when a large set of interdependent variables is used as a model input.