Bad predictions from EO-learn Tree Cover Tutorial

Hello, I am quite new to this so bear with me if I make mistakes or if my description is not specific enough.

I have been trying to work with the following tutorial notebook on some of my own data:

I have succeeded to follow all the steps for the city of Utrecht in the Netherlands, the model runs and also does predictions. However, these predictions are not that great. I noticed in the tutorial that this is also the case however it is not mentioned how or why this is. I found a Medium article using the same tutorial and again the predictions are clearly quite off: Again, this is not mentioned as a problem at all and possible suggestions for improvement are not provided.

Is this the best there is to expect from this or are there things I could do to improve it? I understand that it is not going to be completely right but I expected a bit more. I am trying to understand more of eo-learn for my own project and I’m struggling to find out what I can do to improve or alter these scripts.

1 Like

The best way to improve the results would be to get a better training dataset, or improve the existing one (e.g. by manual labelling).

The data used in the blog post is the EU Tree cover density for 2015 (see here retrieved through Geopedia. The data is outdated (2015), with spatial resolution of 20m (but minimal mapping unit reported is 0.5ha).

@batic Thank you for your reply, I have found new data in the meantime which is probably more suitable. However, I am not quite sure how to implement it in this tutorial. It is a GeoTIFF file, I found the ImportFromTiff EOTask and I suppose this needs to be used but I can’t find an example of implementation. Do you have some pointers?

There is an example on how to use ImportFromTiff in the examples/super-resolution-fastai/DataPrep.ipynb link.

Make sure your tiff is in the same coordinate reference system as the eopatch (I suggest you use gdal to transform if needed).

The documentation on ImportFromTiff is here.

I am also interested in this example, but found no way to improve the prediction results and to do the experiment in the different regions which are out side of EU? Any suggestion would appreciated.