Tutorials + Use Cases

[jejjohnson]

Here is a mega-list of use cases and tutorials we can do to populate the JupyterBook. We can break them down into specific sections. We can break these tutorials down into 4 levels of granularity:

0. Landing Page
1. High-Level Usage
2. Low-Level Usage
3. Contributing

---

Level 0 - Landing Page

This is where the users will enter into the OceanBench package. We will:

1. Describe what is OceanBench
2. Demonstrate how can one install it
3. Describe how does one get access to the data, 4) showcase and how

• What is OceanBench? (Diagram, few paragraphs from paper)
• How to Install
• How to download the data, little bit of dvc fluency
• Getting Started
  • Task -> ML
  • Torch Integration of XRPatcher
  • Map -> Leaderboard
• Showcase 4 Tasks & 4 LeaderBoards
• Upgrading your OceanBench Fluency

---

Level I - High-Level Usage

In this section of tutorials, we look at some "out-of-the-box" solutions that OceanBench provides. This can be useful for people interested in piping preprocessed data for

Tutorials

• Result (Map) -> Leaderboard
• Task 2 XRPatcher

Bonus Tutorials

• Reconstruct Sea Surface Currents from Altimetry Tracks
• Forecasting
• Learning a Latent Space
• Super Resolution, Downscaling, Upsampling
• Denoising

Use Cases

These are self-contained, reproducible use cases which use these out-of-the-box solutions which demonstrate how we can do SSH reconstructions.

• Optimal Interpolation Example w/ AlongTrack Data
• 4DVarNet Starter Example w/ Gridded Data
• Neural Field Example w/ AlongTrack Data

---

Level II - Low-Level Usage

1. Creating Your Own - Hydra Fluency

• Simple Custom Pipeline - preprocessing, postprocessing, plotting

2. Using preconfigured elements of OceanBench

• fetching, reusing
• Load the formatted data
  • AlongTrack
  • Gridded Data
• Load my result with specific postprocessing - reference grid, resampling, evaluation period, domain, coordinate transform
• Create an Evaluation dataset
• Evaluation Maps
• Plotting Maps
• Computing Metrics - PSD Score, nRMSE
• Apply Plotting Configuration - Real Space
• Spectral Space

Modifying Existing

• Create a new Task
  • New Data - changing inputs and/or outputs
  • Changing the domain - Med
  • Changing the evaluation Period - summer
• Create a new metric - e.g. RMSE, (field, field -> scalar)
• adding an additional processing step
  • Preprocessing - e.g. different target grid for XRPatcher
  • Plotting - e.g. filtering

Use Cases

• Spe

---

Level III - Contribution

How do you contribute? How do you add resuable blocks that are resuable by everyone? (package Conventions)

• Add a new metric, e.g. wavelet decomposition integration
• Add a new graphic, e.g. fancy strain histogram plot
• Add drifter data (new data format)
• New XRPatcher for different, e.g. SWOT Calibration
• Global Custom Metrics, e.g. scale-wise, region-specific/dependent (OSE Global DC)

---

---

Pre-Processing

Everything related to accessing available datasets

• Quick dvc tutorial, accessing all of the available datasets (summary table of available ones), downloading, quick plot .

XRPatcher Integration

Some tutorials for how to use XRPatcher and how to integrate it into some other frameworks.

Generic Formulations

• 1D time series (time) + weighted reconstruction
• 2D patches (lat, lon) + weighted reconstruction
• 3D Field (lat, lon, depth/height)

PyTorch Integration

• Simple Usage - Gridded
• Simple Usage - AlongTrack + Concat Multiple Datasets

---

Hydra Usage

Full Walkthrough

Go through each part step by step

• Pipe
• Cross Referencing
• Overriding
• Scripts

0-100 with Hydra

This is how to extend OceanBench with new stuff.

• Changing something in the preprocessing (e.g. coordinate-based)
• Adding a new metric
• Adding a new plot
• Adding a new method
• Adding a new task

Preprocessing/GeoProcessing Pipeline

• Open 1 file + Do 1 pipeline (xr.open_dataset)
• Open Multiple files + Do 1 pipeline (xr.open_mfdataset, preprocess=)
• Multiple files + Multiple Pipelines (xr.open_mfdataset, preprocess=)

Evaluation Pipeline

• Metrics - reference file
• Metrics - reference + comparison file

Visualization

• Plots - Single Variable (SSH map)
• Plots - Multiple Variables (PSD)

LeaderBoard

• Data Challenge 2020a - Gulfstream - OSSE - NEMO
• Data Challenge 2021a - Gulfstream - OSE - Altimetry

Evaluation Metrics

Some standard statistics that one typically uses when evaluating fields for SSH. We typically have two types: 1) gridded and along track. We also need some standard preprocessing steps for some of the metrics

• Regridded
• LatLon degrees -> meters
• Time -> days (or wtv unit)

Gridded

• Skill Scores
• Power Spectrum + Score (Isotropic)
• Power Spectrum + Score (SpaceTime)

AlongTrack

• Skill Scores
• Power Spectrum + Score

---

Visualizations

Showcase some staple visualisations for evaluations

• Using Hydra for Custom Visualisations
  • Temporal Statistics (e.g. nrmse)
  • Maps, e.g. SSA, SLA, Kinetic Energy
  • Power Spectrum (Isotropic)
  • Power Spectrum Score (Isotropic)
• Temporal Statistics (nRMSE, Energy, Enstrophy)
• Derived Variables
  • SSH, SLA, u, v, kinetic energy, relative vorticity, absolute vorticity, enstrophy, strain, Okubo-Weiss
• Power Spectrum + Score (Isotropic)
• Power Spectrum + Score (SpaceTime)
• Animated Maps (gifs + videos - movie)
• Animated Maps (hvplot, geoviews)

---

Data Challenges

A quick rundown and specific details related to the data challenges. It should include:

• Problem settings and experimental setup
• How to access the data via the registry
• A simple preprocessing routine (alongtrack, grid)
• An evaluation script for a demo dataset (DUACS, 4DVarNet)

• 2020a - Gulfstream - OSSE - NEMO
• 2021a - Gulfstream - OSE
• 2022a - GulfStream - OSSE - QG
• 2023a - Mediterranean - OSE
• 2023b - Global - OSE

---

Machine Learning Demos

Some start to finish examples of how OceanBench can be used. Some example applications include

• Sea Surface Height Interpolation
• Forecasting
• Dimension Reduction / Reconstruction
• Surrogate Modeling

SSH Interpolation

Some simple demonstrations from the data challenge

• 4dvarnet OSSE interpolation (@quentinf00)
• 4DVarNet + Transfer Learning OSE Interpolation (@quentinf00)
• nerf OSSE interpolation (@jejjohnson)
• nerf OSE interpolation (@jejjohnson)
• Optimal Interpolation OSE interpolation + sampling (JAX) (@jejjohnson)
• Optimal Interpolation OSE interpolation + sampling (PyTorch) (@quentinf00)