Connecting Earth observations, ground environmental measurements, and high-throughput biodiversity data with an innovative joint species distribution model

Abstract:

In this study, we analyze and predict biodiversity patterns of a landscape-level forest and its surroundings using the combination of high-throughput DNA sequences of species, remote-sensing data of Earth observations, and traditional ground-based environmental data. We take advantage of the progress of artificial intelligence to analyze our data (1186 species over 96 sampling points) with a joint-species distribution model called sjSDM. We test the analytical and predictive power of the model, extracting important environmental factors and testing the model extrapolation. Our final goal is to predict biodiversity patterns in the neighboring area based on the model with existing sampling data and the environmental data of the new area. We collected the remote-sensing data from both airborne Lidar and satellites and our invertebrate samples with pitfall traps in an experimental forest in Oregon, US, extracting DNA with metabarcoding and shotgun sequencing. A preliminary model with a subset of the data shows that the sjSDM model with the environmental variables that we collected can explain 40% of the biodiversity variance of the sequencing data. Among all models that we tried, with different sets of environmental variables, spatial coordinates and human disturbance are most important. The effects of temperature, canopy height, and elevation can be replaced by remote-sensing data. Our preliminary result extracts reasonable environmental factors and explains a considerable amount of variance for biodiversity. Therefore we may predict species patterns of new, neighboring areas successfully which will be shown in the conference.