Google4Habitat: Semiautomatic classification in Google Earth Engine

Background

Protecting habitats requires up-to-date maps, for example for the Red List of Carinthia and the Natura 2000 network. Field surveys are time-consuming, especially in hard-to-reach high mountain areas. The Google4Habitat project therefore developed a method that classifies habitats from freely available satellite data and also reveals changes over long periods. It was tested in Seebachtal in the Hohe Tauern National Park, on around 52 km² between 1,200 and 3,361 m above sea level.

My work

In the team of Gregory Egger (Naturraumplanung Egger), I was responsible for the technical implementation of the classification model:

  • Programming the model in Google Earth Engine
  • Integration of the satellite data: 49 Sentinel-2 and 22 Landsat 8 images from 2022, Landsat 5 from 1992 for the comparison over time
  • Additional data layers from LiDAR: slope, surface roughness and vegetation height, plus the indices NDVI, NDWI and NDSI
  • Classification of 14 habitat types with Random Forest and accuracy assessment at reference points

Methods & results

Reference points were defined for each habitat type based on orthophotos. Two thirds of them were used to train the model, one third to check it. Forests, rock and glaciers are detected reliably. Small habitats such as fens and still water can only be captured to a limited extent with Landsat.

The comparison of 1992 and 2022 shows that little has changed on around 80% of the area. On the remaining 20%, there were shifts between forest and open land, partly due to avalanches and debris flows.

Egger, G., Preinstorfer, S., Kollmann, M., Becker, I., Izquierdo-Verdiguier, E. & Paul, M. (2024): Google4Habitat – a novel method for remote sensing-based habitat classification using Google Earth Engine. Carinthia Nature Tech 1(1): 8–28.

More about this service: Remote sensing and satellite image analysis