Skip to content
Snippets Groups Projects
Select Git revision
  • main
1 result

mgan

  • After an exchange of emails initiated by Dr. Jordi Inglada, I carried out a careful evaluation of the whole data processing pipeline of our model and of the HLS documentation, coming to the conclusion that our method provides , while HLS, despite being Analysis Ready Data (ARD), needs to be cleaned (outliers removed) before feeding them into ML/DL models. I also utilized some box plots to explain this.

    As originally I computed the metrics for HLS on uncleaned data, I carried out a new calculation for a fair and comprehensive evaluation with the ARD data produced by our method. I hereby attach the table with the distance between cleaned HLS and real data, which are indeed better than those reported in the table on the journal paper on uncleaned HLS data.

    I hereby attach the table with the distance between cleaned HLS and real data, which show that if outliers are removed from the HLS data before evaluating the metrics, the distance between HLS and MGAN (our method) is reduced significantly.

    Considering these find, the following eq. holds as of spectral agreement of the products with respect to the target L8: S2 < HLS <(=) MGAN.

    Data Access and Preparation The S2 tile considered in the At the time of the publication of the paper, the HLS data used were not publicly available on the HLS website, and were generated and kindly made available by Brian M. Freitag and Manil Maskey(NASA). The S2 data were retrieved through the Copernicus hub, while the L8 data thorugh EarthExplorer. We converted the data to HDF5 format for parallel I/O on HPC machines (these files can be made available upon request).