Using k-means clustering as a robust and repeatable method to determine representative climate/eco-regions

The Amazon has a significant role in regional and global climate and carbon cycles. Having a robust method to determine and understand representative climatological sub-regions, and being able to sample these regions would help improve modelling the land surface and carbon cycle. The regions would also inform understanding of the representativeness of in-situ observations thereby correctly interpolating and bias correcting results as well as informing locations of future sites. Here the regions are calculated using a machine learning algorithm called k-means clustering, which groups datapoints which are close in variable space, hence have similar climatological characteristics. Various combinations of input variables and number of clusters (k values) were explored but the final results used annual average temperature, annual average precipitation and soil phosphorus as inputs, which produced contiguous regions which were easy to interpret. These regions were then evaluated using marginal distributions of the input variales and by exploring the above-ground-biomass distribution. This was performed for both present day observational data inputs and for projected climate data using ISI-MIP bias correct climate projections, exploring how the regions may change in future climates. This showcases the Amazon as an example, but more importantly highlights a robust technique for determining eco-regions which can be applied to different locations and climate scenarios.