search

dll-ncai / AI-ForestWatch

MIT License
8 stars 7 forks source link

Clash between paper and GitHub #2

Closed srinathjayaraman closed 2 years ago

srinathjayaraman commented 2 years ago

Hello,

I am a masters student at TU Delft, Netherlands. I am replicating this paper as part of my masters thesis. I noticed a difference between what the original paper says vs. what the GitHub README says regarding UNet training and testing. I have highlighted the difference below (emphasis mine).

Quote from the paper - A UNet model is trained with an 18-channel input image and the generated digitized forest cover map of year 2015 as ground truth. After a model is trained successfully, inference is performed on the clean Landsat-8 district images of years 2014 to 2020.

Quote from the GitHub README - We used images from 2014, 2016-2020 for training and 2015 for inference.

I spoke to one of the original authors of the paper, Mr. Annus Zulfiqar, and he clarified that the original paper contains the right information.

Images from 2015 were used as ground truth for training and testing (80/20 split). Real world performance of the UNet was then tested on the images from 2014 and 2016-2020.

I'm opening this issue so the contradiction can be corrected and further confusion will be prevented.

Thank you!

mohsin432 commented 2 years ago

Dear @srinathjayaraman Thank you for mentioning the error. The information is updated.