Deep Learning Methods for Extracting Habitat Summaries from Remotely Sensed Data for Species Distribution Modeling

Methods for collecting environmental variables for species distribution models (SDMs) generally consist of field-based methods or statistics of remotely sensed data. These statistics (e.g., mean and standard deviation of image bands) may not be sufficient for capturing complex land cover patterns and are rudimentary when compared to state-of-the-art computer vision techniques. Oversimplified representations of landscapes may prevent models from determining the impacts of fragmentation on species distributions. Given the advancements in computer vision due to deep learning, features extracted by deep neural networks have the potential to characterize habitats better than methods currently used to summarize remotely sensed data. More descriptive habitat variables are likely to lead to more informative SDMs.

In this work, we trained deep neural networks on a variety of tasks (e.g., to classify land cover from aerial images) to obtain models that can compute habitat features from remotely sensed images; the habitat features can then act as inputs to any style SDM. We compared the habitat features computed from deep networks to several sets of habitat features commonly used in SDM (e.g., statistics of remotely sensed data) by modeling bird occurrences in the state of Oregon. We modeled five species with Occupancy-Detection and Random Forest models using data from eBird.

Surprisingly, we have found little difference in model performance when predicting species occurrences with simple summary statistics versus habitat features computed from the deep networks. We will discuss several hypotheses to explain these results and promising directions for deep neural networks to provide informative habitat features.