mayrajeo / lulc_ml

Codebase for land cover classification with U-Net
MIT License
18 stars 6 forks source link

Problem with classifications #2

Open matiasmsm opened 10 months ago

matiasmsm commented 10 months ago

Hi!

I hope you are doing well and thank you for publishing this project.

I'm having problems classifying Sentinel-2 images taken inside the territory of Chile. The model consistently predicts water bodies where there's vegetation and other misplaced clases. ¿Any idea whats going on?

Thank you!

Screen Shot 2024-01-08 at 17 14 52 Screen Shot 2024-01-08 at 17 16 30 Screen Shot 2024-01-08 at 17 20 47 Screen Shot 2024-01-08 at 17 25 50 Screen Shot 2024-01-08 at 17 27 48

rongtongxueya commented 10 months ago

hello, I'm curious, did you cut the original tif images into tiny pieces like this? Is it similar to becoming the same as RGB natural images? Is that how it goes into the network

matiasmsm commented 9 months ago

Hi! Yes, i did divide the image into smaller ones. I think the problem i'm having arises when i'm normalizing the data. From the paper and the documentation it wasn't clear to me how the author did the normalization of the bands and the spectral indices. If you have any idea i would much appreciate it if you could tell me! Cheers.