Interpolation of temporal biodiversity change, loss, and gain across scales: a machine learning approach

1. Estimates of temporal change of biodiversity, and its components loss and gain, are needed at local and geographical scales. However, we lack them because of data in-completeness, heterogeneity, and lack of temporal replication. Hence, we need a tool to integrate heterogeneous data and to account for their incompleteness. 2. We introduce spatiotemporal machine learning interpolation that can estimate cross-scale biodiversity change and its components. The approach naturally captures the expected and complex interactions between scale (grain), geography, data types, and drivers of change. As such it can integrate inventory data from reserves or countries with data from atlases and local survey plots. We present two flavors, both blending tree-based machine learning (random forests, boosted trees) with advances in ecolog-ical scaling: The first combines machine learning with species-area relationships (SAR method), the second with occupancy-area relationships (OAR method). 3. Using simulated data and an empirical example of global mammals and European plants, we show that tree-based machine learning effectively captures temporal biodi-versity change, loss, and gain across a continuum of spatial grains. This can be done despite the lack of time series data (i.e., it does not require temporal replication at sites), temporal biases in the amount of data, and highly uneven sampling area. These estimates can be mapped at any desired spatial resolution. 4. In all, this is a user-friendly and computationally fast approach with minimal require-ments on data format. It can integrate heterogeneous biodiversity data to obtain esti-mates of temporal biodiversity change, loss, and gain, that would otherwise be invisi-ble in the raw data alone.