Thank you for releasing CarbonSense and EcoPerceiver. I am trying to reproduce the experiments reported in the CarbonSense paper, especially the EcoPerceiver results in Table 2, and I noticed that the current code/config seems to include some inputs that are not clearly described in the main paper.
I would like to ask a few clarification questions about the exact setup used for the paper experiments:
-
Input modalities
In the paper, EcoPerceiver is mainly described as using EC/tabular predictors and MODIS spectral data. However, the current code appears to support additional inputs such as PhenoCam, latitude/longitude/elevation, and IGBP embeddings.
Were the results reported in the paper trained with only EC predictors + MODIS, or did they also include PhenoCam, lat/lon/elev, and IGBP as model inputs?
-
Train/validation/test split
Is the train/validation/test site split used in the paper the same as the one currently provided in the config files?
If not, is there a released split file or random seed that should be used to reproduce the Table 2 results?
-
Dataset version
I noticed that there seem to be two dataset versions associated with the project. For reproducing the paper experiments, which dataset should be used?
4.Training configuration
Could you point me to the exact training configuration used for the paper experiments?
It would be very helpful if you could provide the exact config file, commit hash, or command line used to reproduce the results in the paper.
Thank you very much for your time and for making this dataset and code available.
Thank you for releasing CarbonSense and EcoPerceiver. I am trying to reproduce the experiments reported in the CarbonSense paper, especially the EcoPerceiver results in Table 2, and I noticed that the current code/config seems to include some inputs that are not clearly described in the main paper.
I would like to ask a few clarification questions about the exact setup used for the paper experiments:
Input modalities
In the paper, EcoPerceiver is mainly described as using EC/tabular predictors and MODIS spectral data. However, the current code appears to support additional inputs such as PhenoCam, latitude/longitude/elevation, and IGBP embeddings.
Were the results reported in the paper trained with only EC predictors + MODIS, or did they also include PhenoCam, lat/lon/elev, and IGBP as model inputs?
Train/validation/test split
Is the train/validation/test site split used in the paper the same as the one currently provided in the config files?
If not, is there a released split file or random seed that should be used to reproduce the Table 2 results?
Dataset version
I noticed that there seem to be two dataset versions associated with the project. For reproducing the paper experiments, which dataset should be used?
4.Training configuration
Could you point me to the exact training configuration used for the paper experiments?
It would be very helpful if you could provide the exact config file, commit hash, or command line used to reproduce the results in the paper.
Thank you very much for your time and for making this dataset and code available.