Problem in replicating the Results

[ghost]

Dear Orion AI-Lab,
I am stuck in replicating the results as reported in the paper while running the PAD models (UNET, CONV-LSTM and CONVSTAR).
For running the models I used the following command
Patch file name: "2019_31TCJ_patch_24_03.nc"
In case of Unet:
python visualize_predictions.py --model unet --root_path_coco coco_files/ --netcdf_path dataset/netcdf --bands B02 B03 B04 B08 --img_size 61 61 --requires_norm --load_checkpoint weights/unet/epoch=16-step=28695.ckpt
In case of Conv-LSTM:
python visualize_predictions.py --model convlstm --root_path_coco coco_files/ --netcdf_path dataset/netcdf --bands B02 B03 B04 B08 --img_size 61 61 --requires_norm --load_checkpoint weights/conv_lstm/epoch=17-step=60749.ckpt
In case of Conv-star
python visualize_predictions.py --model convstar --root_path_coco coco_files/ --netcdf_path dataset/netcdf --bands B02 B03 B04 B08 --img_size 61 61 --requires_norm --load_checkpoint weights/conv_star/epoch=15-step=27007.ckpt

I am attaching the output file in case of ConvStar.

If you could please assist me in this regard.
Many thanks for taking your time.

Regards,

[paren8esis]

Hello,

In order to isolate the predictions for the parcels you should also add the argument --mask_parcels in the command. Is this what you mean?

[ghost]

Thank you so much for the kind assistance.

In continuation to the discussion above,
By adding argugment --mask_parcels , I used the commands to run the PAD models,
In case of Unet:
python visualize_predictions.py --model unet --root_path_coco coco_files/ --netcdf_path dataset/netcdf --bands B02 B03 B04 B08 --img_size 61 61 --requires_norm --load_checkpoint weights/unet/epoch=16-step=28695.ckpt --mask_parcels
In case of Conv-LSTM:
python visualize_predictions.py --model convlstm --root_path_coco coco_files/ --netcdf_path dataset/netcdf --bands B02 B03 B04 B08 --img_size 61 61 --requires_norm --load_checkpoint weights/conv_lstm/epoch=17-step=60749.ckpt --mask_parcels
In case of Conv-star
python visualize_predictions.py --model convstar --root_path_coco coco_files/ --netcdf_path dataset/netcdf --bands B02 B03 B04 B08 --img_size 61 61 --requires_norm --load_checkpoint weights/conv_star/epoch=15-step=27007.ckpt --mask_parcels

I am attaching the output file in case of ConvStar.

If you could assist in this matter.
Regards,

[paren8esis]

Well, it seems that the pretrained model has a hard time classifying that particular patch. Have you tried others and observed the same pattern?

We have to note here that all PAD models showed a lot of confusion regarding the "Oats" class and tended to misclassify other crops as oats. You can also see that in the confusion matrices in Fig. 11 of the article. And I can see that same problem in the patch you have plotted.