Using Taxonomically-Informed Convolutional Neural Networks To Predict Plant Biodiversity Across California From High-Resolution Satellite Imagery Data

Abstract:

In order to understand how spatial distribution of species across ecological communities will shift under global environmental change, we first need robust, fine-scaled models of how species are currently distributed spatially. Yet, predicting the spatial distribution of multiple species on a fine-grained scale from remote sensing data is still an open problem in landscape ecology. California’s unique ecological makeup makes it an optimal system to investigate correlation of biodiversity with remote sensing data. However, standard species distribution models and plant biodiversity maps for California predict solely on interpolated environmental variables and rely on automatically generated negative sampling techniques that have known biases. Here, we propose learning a presence-only joint species representation from both remote sensing data and interpolated environmental variables. Specifically, we train a convolutional neural network on high-resolution satellite imagery data from NAIP and interpolated environmental variables from WorldClim paired with geolocated observations from GBIF to jointly predict plant species presence across California. To aid in predicting over thousands of different classes, we introduce a novel taxonomically-informed loss, where species, genus, and family are classified for each training example. Our model predicts at a higher precision and recall than the state-of-the-art multi-response Random Forest model. We also introduce a novel plant biodiversity map of California at a finer-grained resolution than current models. Analyzing poorly predicted regions of California reveals areas of low data sampling and endemic habitats, and pinpoints microclimates for future observation collection. Our model highlights the power of incorporating remote sensing data into species distribution modeling for recovering finer-scale interactions for both biodiversity mapping and species presence prediction under current and future climates.