Deep Learning Satellite Imagery Land Cover Classification with U-Net

[@ahmedameen4587]

Good work❤❤

[@explorethebeautyofnature3530]

Thank you for the excellent presentation. I have a question about the methodology: If the U-Net model requires training data from a completed pixel-based classification, and its predictions are inherently limited by the accuracy of that initial classification, what advantages does this two-step approach offer over directly using pixel-based classification methods? I am particularly interested in understanding how this architecture compensates for or improves upon the initial classification's potential limitations.

[@ramiqcom]

The sampe is not from a pixel based classification. It was vector based which I rasterized. The advantage is that it could understand the spatial pattern instead of justice the reflectance pattern. Like understanding forest density class based on its nearby pixels instead of its own.

[@explorethebeautyofnature3530]

@@ramiqcom Thank you for the clarification. I'll be glad to learn about how the manual labeling was done to cover the entire Landsat image (without gaps/overlaps), and how long it took

[@ramiqcom]

@@explorethebeautyofnature3530 it just classical digitization/delineation. You can search for it.

[@sairam7801]

can the same preprocess be applied to high resolution aerial imagery if i need to extract classes for U-net based landcover classification? and does the same process of code be modified to process aerial imagery land cover classification? Ofc its a RGB imagery as a 4 band raster. Thank you

[@ramiqcom]

Yes you could, you can use any imagery as long as you have the sample

[@dhirojkumarbehera3293]

Good work...just a question that while applying for the whole image it's showing error of graph execution as the model is trained for 128*128 while predicting for the whole image of larger size, can you let me know the possible solutions plz...

[@ramiqcom]

Make sure the ratio is the same. So 128 x 128, 256 x 256, 1024 x 1024. If you use let say 1024 x 928, it will not work.

[@dhirojkumarbehera3293]

@ramiqcom I am using the same ratio for but it's for the whole image not for the patches so it's showing error ...I hope the model is trained for 128*128 so it will show same problem...but in your code how u ran for whole image then?

[@ramiqcom]

Yes it will work for the whole image too although training patch is 128x128. The whole image I used when I read it, I make the ratio to be 2560 x 2560 which is the same ratio. This can be predicted by model.

[@dhirojkumarbehera3293]

​@@ramiqcomthank you for the suggestions...still it's showing the same error what is the code modification we need to do? Please explain...

[@ramiqcom]

When you read the image what is the shape?

[@inspirationalknowledge3609]

how can i access the code

[@ramiqcom]

Read the desc

[@adibsulthonmuammal9544]

berarti ini CNN2D?

[@ramiqcom]

yep

[@EsraaTarek-f7y]

I have Arcgis pro 3.0 will it work with it?

[@ramiqcom]

I dont know this is for python

[@inspirationalknowledge3609]

Can’t access the code in gitup

[@ramiqcom]

What did you get?

[@Ramilacookware]

🎉❤❤❤❤❤❤❤

[@Ramilacookware]

Do you know difference between Index formula and regression formula for calculating pollution? 😂❤

[@ramiqcom]

I dont know, I have not try any

[@inspirationalknowledge3609]

when i try to access the code at github it says data is too larged how can i resolve this issue? can i get your personal email i want to discuss more please

[@ramiqcom]

What data too large?

[@inspirationalknowledge3609]

@@ramiqcom can i get your email address i have some queries to discuss