marine-debris / marine-debris.github.io Goto Github PK

I am trying to replicate the Random Forest model that uses Spectral Signatures and Spectral Indices, but I am unable to achieve the same recall (Pixel Accuracy) results.

I noticed that the results can change between Python versions depending on the packages, so I am using the same environment provided by your yml.
My hyper-parameters are the same (including seed), I am also using the weights calculated from Confidence and Water Super Class. I am using the provided dataset.h5 and dataset_si.h5.

These are my results:

As you can see, recall is slightly different from your PA results:
1: 0.92; 2: 0.93; 3: 0.92; 4: 0.27; 5: 0.7; 6: 0.82; 7: 0.83; 8: 1; 9: 0.48; 10: 0.83; 11: 0.33

Do you know what can be causing this, since the seeds and data are all the same?

Thank you!

Hi, why MARIDA algorithm uses l_redge instead of l_red proposed by Biermann et al. 2020 to calculate FDI?

r_acc = band6 + 10*(band10 - band6)*(l_nir - l_redge)/(l_swir - l_redge)

Hello everyone,

The link for the dataset_si.h5 and dataset_glcm.h5 is broken :(((

Could someone please fix the link?)
I was trying to run on my own pc, but the progress bar stops moving at a point.

I would be very thankfull if someone can share these files with me.

The indices/ and texture/ folders as well as the dataset_si.h5 and dataset_glcm.h5 files from here.

Hi, I'd like to know whether all the hyperparameters in this paragraph from your paper were chosen with a grid search.

During the U-Net training process, the Adam algorithm was employed to minimize the Cross-Entropy loss with an initial learning rate of 2x10-4. Moreover, we utilized early stopping based on the loss of the validation set and trained for 44 epochs. After the 40th epoch, the learning rate was reduced to 2x10-5. The selected batch size was 5 samples. We also employed random rotations of the input images by -90˚, 0˚, 90˚, or 180˚ and horizontal flips in order to augment the dataset. The selection of the hyperparameters above and training set-up was based on grid search in the validation set

Could you tell me any other batch sizes and learning rates that you used? I want to experiment with larger values for these and wanted to know if you have done that already and what your results were. Also, did you try using SGD with momentum, another learning rate scheduler, or additional augmentations?

Basically, I want to know a bit more details about your hyperparameter search. Thank you.

Would you please tell me the exact way of installing it?
where should I move the downloaded files?
I run the script (conda env create -f environment.yml) on anacondas prompt but I received an error.

Your help is appreciated

Hi,
I am using MARIDA to perform a random forest classification. I have a doubt about the atmospheric correction to apply in ACOLITE to the S2 tiles before performing the classification. As stated in Kikaki et al. (2022) rayleigh reflectance values are extracted, while in Kikaki (2020) surface reflectance values are obtained.
Could you please confirm exactly which ACOLITE's outputs should I obtain to obtain the best classification results? rhos_: ρs, the surface level reflectance or rhorc_: ρrc, the Rayleigh corrected reflectance ?

Thank you for you support.
Andrea