Description: This QGIS project visualises preliminary model inference results of a binary habitat classification. They have been obtained by running a pixel-wise, multitemporal time series deep learning model on Sentinel-2 data. Specifically, a fully convolutional network (FCN) has been trained on a 3-year stack of monthly cloudfree Sentinel-2 data containing all 10 and 20m bands from the vegetation periods 2019-2021. The model doesn't differentiate between different habitat types but performs a binary classification habitat vs. non-habitat (e.g. urban areas, arable land, etc). The inference results are calculated for validation data that have been used during training only for early stopping. Continuous spatial predictions are obtained by merging the pixel-wise prediction results in the spatial dimension. Specifically, the following layers are available: ├── `ground_truth_vector.geojson` -> ground truth habitat data compiled from different mapping in-situ campaigns (all types 2010 & grassland mapping 2020) ├── `ground_truth_raster.tif` -> ground truth habitat data compiled from different mapping in-situ campaigns (all types 2010 & grassland mapping 2020) ├── `predictions_plain.tif` -> model predictions from a 1st model version - predicted habitat probabilities ├── `predictions_plain_binary.tif` -> model predictions from a 1st model version - binary predictions (habitat vs. non-habitat) ├── `predictions_plain_flaws.tif` -> model predictions from a 1st model version - errors in comparison to ground truth ├── `predictions_improved.tif` -> model predictions from a 2nd model version - predicted habitat probabilities ├── `predictions_improved_binary.tif` -> model predictions from a 2nd model version - binary predictions (habitat vs. non-habitat) └── `predictions_improved_flaws.tif` -> model predictions from a 2nd model version - errors in comparison to ground truth The layers `*_flaws.tif` are encoded as follows: * 1 - false negatives -> habitats that exist in reality are not recognized * 2 - false positives -> habitats are predicted at locations where there are no habitats in reality
Global identifier:
Other(
"991303740715401216",
)
Tags: Grünland ? Habitat ? Ackerfläche ? Vegetationsperiode ? Daten ? Klassifikation ? Urbaner Raum ? Zeitreihe ? Biotoptyp ?
Bounding boxes: 6.4253° .. 7.06878° x 50.08101° .. 50.37594°
License: Creative Commons Namensnennung 4.0
Language: Englisch/English
Issued: 2026-06-09
Modified: 2026-06-09
Last harvest: 10.08.2026 23:19
Update frequency: not planned
Accessed 11 times.