Deep learning meets Sentinel imagery to predict species distributions.

Biodiversity is being lost at an unprecedented rate worldwide. To better address this crisis, efficient methods for monitoring biodiversity are essential. At the same time, Earth-observation satellites provide vast amounts of data that can be leveraged with the help of artificial intelligence. Here, we present a pipeline that uses Sentinel-2 Level-2A imagery and a convolutional neural network (CNN) to predict the potential distribution of individual species.

Our approach ingests all 12 spectral bands from multiple unprocessed images of the same location. The workflow begins with a dataset containing presence–absence records and their coordinates for the species of interest, along with an optional definition of the target prediction area. The pipeline then automatically downloads the required satellite imagery products. For each presence–absence record, we use satellite images from the corresponding year of data collection. To train the CNN model, we extract an 18 x 18-pixel neighbourhood centered on each record location. The pipeline then trains the CNN and generates probability-of-occurrence predictions. Predictions are provided as georeferenced TIFF images, with probabilities computed for each 10 x 10 m pixel.

Two of the major threats to natural ecosystems are the spread of invasive species and the extinction of rare endemics. Therefore, we demonstrate the utility of our approach using two example species: an invasive mosquito and an endemic plant. The Asian tiger mosquito (Aedes albopictus), a recently established species in Europe, can act as a vector for several pathogens (e.g., yellow fever). Training data for this species were obtained from the citizen-science programme szunyogmonitor.hu. Our second case study focuses on Seseli leucospermum (“magyar gurgolya”), a strictly protected plant species endemic to Hungary, for which records were obtained from the biodiversity databases of Hungarian national parks. We successfully trained our model for both species, achieving validation accuracies above 80% and now calculate the predictions. Future work will include field validation of the predictions.

This study was supported by the Hungarian Research, Development and Innovation Office (grant K138503).