sjSDM

sjSDM fits joint species distribution models (jSDMs): models of how a whole community of species responds to the environment, to space and to each other. The difficult part of a jSDM is the species–species covariance matrix, which most implementations approximate with a small number of latent variables. sjSDM estimates the full matrix instead. It remains fast enough for large community datasets, including eDNA data, because it approximates the joint likelihood by Monte Carlo integration, implemented in PyTorch and run on CPUs or GPUs.

Beyond fitting, the package partitions explained variation into environmental, spatial and biotic components and can break this down by species and site to reveal the internal structure of a metacommunity. It offers several response distributions, including the classical multivariate probit model, terms for spatial autocorrelation, and the option of replacing the linear predictor with a deep neural network.

Authors: Maximilian Pichler (maintainer), Florian Hartig. Contributors: Wang Cai.