ERODYN

  The study area is located in the Funes valley in South Tyrol, Italy, where shallow erosion plots have been increasing in recent years and decades. The study area lies at an altitude of more than 2300 m above sea level. Using different techniques, we mapped the grassland vegetation in this area at three different scales: species level, grassland community level and vegetation region level (local scale). The main source of data was a hyperspectral survey with a total of 28 spectral bands (506 nm to 896 nm), acquired with a spatial resolution of 5 cm, based on UAV. Our ground reference database consisted of detailed ground measurements within 50×50 cm quadrats. In total, we made field spectroradiometer measurements covering the spectral range from 339 nm to 2500 nm (1024 spectral bands), ground hyperspectral measurements, and sampled different grassland communities within the quadrats.

  As a result of the data acquisition, the integration of ground and aerial data, the study was able to determine the vegetation cover of the area at the grassland community level. Furthermore, their content was quantified using a standardised sampling method – thus a detailed species list and database was built up. By surveying the grassland communities, a distribution map of the communities was produced using machine learning based pixel-based classification. Based on this, a small-scale vegetation region map (interpreted at local scale) was computed using an object-based image analysis (OBIA) approach (including multi-resolution segmentation) – Thus generalising and typifying the land cover of the area to support the further work to determine the most endangered regions.

  The overall accuracy achieved was 75.57% and the average user accuracy was 79.08% according to the confusion matrix validation. The false positive results were not significant, with a lower incidence of misclassification. The main challenge was the similarities in spectral signature between overlapping species and classes.