Automatic segmentation for insular habitats along the Loire River: test in a National Nature Reserve, Mareau-aux-Prés France
Résumé
Monitoring vegetation dynamics in nature reserves requires accurate mapping of habitats. In August 2017, we used a drone to acquire LAS images and high-resolution (2cm)orthophotos on islands of nature reserve of St-Mesmin along the Loire river. At the same time, we mapped habitats based on vegetation sampling every 15 m and the phytosociological nomenclature published by the Botanical Conservatory of the Paris Basin. The OTB workflow allowed us to map 9 vegetation types but with an overall kappa of 0.50 only. If the results were satisfactory for the different riparian forest types (min F-score = 0.70) , the various tools tested under OTB failed to correctly classify open areas as reed-bed (F-score = 0.31) and grassland (F-score = 0.18). Two hypotheses can explain this problem: the low accuracy of the field GPS data and some subjectivity in the interpretation of the phytosociological nomenclature.