Mapping Fire Salamander (Salamandra salamandra) Habitat Suitability in Baden-Württemberg with Multi-Temporal Sentinel-1 and Sentinel-2 Imagery

English. Remote sensing image classification is used in land cover, forest type and tree species classifications but rarely considered for habitat suitability modelling of animal and plant species. It is instead common that land cover products derived from remote sensing data are used in these modelling problems, even though satellite imagery can provide more detailed information. The aim of this project was thus to explore remote sensing image classification methods to classify land covers in Baden-Württemberg based on their habitat suitability for the fire salamander (Salamandra salamandra). Fire salamanders depend on both suitable aquatic and terrestrial environments, and the classification was therefore applied on multi-temporal Sentinel-1 and Sentinel-2 images combined with a waterway proximity layer derived from OpenStreetMap data. The classification used a random forest classifier which was trained to discriminate between positive samples from tree covered areas within 300 m of fire salamander observations and unlabelled samples drawn from a regular grid with 1500 m point spacing in the study area. Two classification methods were evaluated: pixel-based, in which single pixels are used in the classification, and superpixel-based, in which the classification was performed on mean pixel values of approximately equally sized (∼1 ha) vectorized regions of similar pixels derived from a Simple Linear Iterative Clustering (SLIC) segmentation of the study area. The resultant classifications were compared against a model with land cover data from Copernicus Land Monitoring Service, and the evaluation showed that the image classifications were able to discriminate better between positive and unlabelled test samples. The superpixel-based classification further achieved a higher evaluation score (AUC: 0.91) than the pixel-based classification (AUC: 0.90), and was thus the best model in the analysis. An exploratory analysis of the predictions based on LUCAS 2018 survey points further indicated that the models predicted high fire salamander habitat suitability in tree covered areas situated within roughly 200 m of stream and river features, with more than 10 % canopy cover, and with more than 25 % of broadleaved trees in the canopy composition. Remote sensing image classification for fire salamander habitat suitability modelling was concluded applicable at regional mapping scales, and more generally for habitat suitability modelling of species that are highly dependent on land cover characteristics.

English. Is it possible to identify fire salamander habitats from space? That is the question this project tries to answer. The European fire salamander is currently threatened by habitat destruction and the continued spread of a deadly disease caused by a fungus called Bsal. Protecting the fire salamander is thus a hot topic nowadays, and one important tool that can assist conservationists in their efforts to help the species is a habitat suitability map. These maps are usually generated by relating observations of the species with environmental data sets, but in this project, a method based on satellite image classification was used to produce a map of habitat suitability for the fire salamander in the German state of Baden-Württemberg. The image classification was performed on a set of images from the European Sentinel satellites, but also included a layer which described distance to OpenStreetMap waterways – representing potential breeding sites of the species. The technique was based on relating fire salamander observations with physical characteristics in the imagery and thereafter use this relationship to predict habitat suitability over Baden-Württemberg. Two classification methods were tested: “pixel-based”, in which each pixel was treated individually in the classification, and “superpixel-based”, in which the study area was first partitioned into regions of similar pixels, which were then used in the classification by considering the mean pixel value over each region. The pixel-based and the superpixel-based image classifications were compared against a model which was built using available land cover data sets, and the evaluation showed that both image classifications were slightly better at predicting fire salamander habitat suitability than the land cover model. This result indicates that satellite imagery can provide even more useful information to distinguish fire salamander habitats than what common land cover data sets do. The superpixel-based classification also performed better than the pixel-based method, which suggests that region based image classification techniques are suitable to evaluate in other habitat suitability mapping projects as well. Exploration of the final habitat suitability maps further showed that high fire salamander habitat suitability in Baden-Württemberg was found in tree covered areas with broadleaved and mixed tree stands within 200 m of streams and rivers. The results motivates the use of satellite image classification techniques to map habitat suitability for the fire salamander or other species for which detailed land cover information is very important.