Semantic segmentation export polygon about eo-learn HOT 4 CLOSED

Hi @tscavnicar,

There exists a RasterToVector task which vectorizes any discrete raster feature into a new vector feature. The obtained vector feature will be in a form of geopandas.GeoDataFrame which you can easily save as e.g. shapefile.

I hope this is what you were looking for.

from eo-learn.

Hi @AleksMat,
good to have the info.

Can the tree-cover-keras example be used to predict custom objects, e.g. buildings, roads, farm fields?

from eo-learn.

Hi @tscavnicar,

The example uses a U-net architecture which performs pixel-wise classification (i.e. one-label-per-pixel). If the ground-truth labels contain enough examples of what you want to predict (e.g. buildings, crop fields), then the architecture should be able to predict them if it is re-trained on that specific data. If instead you would like to do object detection (i.e. put a bounding box around specific objects), you'd need a different architecture, for instance a YoLo.

Be aware that Sentinel-2 images have a spatial resolution of 10mx10m, so predicting small objects (single houses, secondary roads) will not be possible. Our land-cover classification notebook and blog posts show an example of what can be achieved. The U-net model would replace the gradient boosting machine classifier.

Hope this helps.

from eo-learn.

Hi @devisperessutti,
great for info, it helps.

from eo-learn.