Train DeepSensor on Individual Great Lakes Using Developed Masks

[DaniJonesOcean]

Task Description:

Train the DeepSensor model on data for individual Great Lakes using the specific masks developed for each lake. This task aims to refine and optimize the model performance for localized sensor placement and better understanding of surface temperature variability in each lake.

Checklist:

• Data Preparation:

  • [ ] Retrieve the individual Great Lakes masks developed during the project.
  • [ ] Ensure that the masks are correctly formatted and aligned with the input data requirements for DeepSensor.
  • [ ] For each lake mask, verify that the relevant observational data is available and properly preprocessed.

• Model Training:

  • [ ] Set up the DeepSensor model environment
  • [ ] Configure the model to use the specific mask for each Great Lake (e.g., Lake Superior, Lake Michigan, Lake Huron, Lake Erie, and Lake Ontario).
  • [ ] Train the model separately for each lake, ensuring that the appropriate parameters and settings are used.
  • [ ] Monitor the training process and record key performance metrics for each lake.

• Evaluation and Analysis:

  • [ ] Evaluate the model performance for each lake individually.
  • [ ] Compare the results against lakewide models
  • [ ] Document the findings for each lake.

• Documentation and Reporting:

  • [ ] Create a Jupyter notebook to document the model training process for each lake, including code, results, and visualizations.
  • [ ] Save the trained models and relevant output files.

[eredding02]

@DaniJonesOcean I have graphs of predictions of the ConvNP trained on individual lakes in this slide deck. The majority of the predictions did significantly worse. Some hypotheses as to why this happened are that due to the data-hungry model it needs more years or that it needs the rest of the lakes to contextualize predictions.

[DaniJonesOcean]

Great work, @eredding02! Really interesting results.

I think any further work here will fall under the header of "training refinement"; we could consider making that a separate issue to capture your efforts there.