Developing a joint species distribution model for plant species in Switzerland using remotely sensed data and computer vision

Species Distribution Modelling (SDM) is a vital instrument for predicting the spatial and temporal patterns of biodiversity. By leveraging environmental variables and occurrence data, SDM can forecast the potential distribution of species, pinpoint areas of high biodiversity and habitat suitability, and guide conservation and management strategies. This is particularly relevant in the face of climate change and other human-induced disturbances that can affect biodiversity. In this work, I propose a Deep Learning approach based on the Inception V3 CNN architecture to model the distribution of 4,466 plant species in Switzerland on a yearly basis, by leveraging remote sensing data. This master’s thesis tries to answer the following questions: (i) How does the choice of input features affect the performance of the model? (ii) What level performance can be expected from the model for common species and rare species? (iii) How accurate are the species distribution predictions for the years that are not included in the training set? (iv) How do the biodiversity metrics derived from the model’s results compare to the ones from a reference database? I found that the bioclimatic variables with a temporal resolution of one year are not suitable for species distribution modelling. The resulting model has a median AUC value of 0.88 for both rare and common species, however, the variability of AUC for rare species is higher. The ability of the model to predict species distributions for the years that it was not trained on is slightly lower than for the years occurring in the training set, and the biodiversity metrics derived from the model predictions match the patterns observed in reference datasets for some areas in Switzerland, but their correlation coefficient is 0.06.