Satellite Embedding-Based Deep Learning Framework for Reconstruction of Subsurface Ocean Temperature from Surface Satellite Observations
SIH 26066 – OceanEmbed: An AI-based oceanographic framework that uses multi-source satellite surface observations to accurately reconstruct depth-wise subsurface ocean temperature for the North Indian Ocean, with initial high-resolution demonstration focused on the Bay of Bengal.
-
No Hardcoded Values: All temperature predictions, predictive uncertainty bounds, accuracy metrics, RMSE, Bias, Pearson correlation (
$r$ ), and error percentages are computed dynamically from actual datasets and the trained deep neural network. -
Clear Indicators: When uninitialized or running in simulation demo mode, the application explicitly tags fields as
DEMO DATA – NOT SCIENTIFIC PREDICTIONorModel not trained.
-
Primary Demonstration Region: Bay of Bengal (
$5^\circ\text{N}\text{ to }22^\circ\text{N}, 80^\circ\text{E}\text{ to }100^\circ\text{E}$ ) -
Expandable Region: North Indian Ocean (
$5^\circ\text{N}\text{ to }30^\circ\text{N}, 45^\circ\text{E}\text{ to }105^\circ\text{E}$ ) -
Spatial Resolution:
$0.25^\circ \times 0.25^\circ$ - Temporal Resolution: Daily
- NOAA OISST: Sea Surface Temperature (SST) and error field.
- Sea Surface Salinity (SSS): Halocline / salinity surface layers and river runoff plumes.
- SSH / SLA: Sea Surface Height / Sea Level Anomaly (capturing cyclonic & anticyclonic mesoscale eddy pumping).
-
Surface Currents: Geostrophic + Total Currents (
$U_{\text{curr}}, V_{\text{curr}}$ ). -
Surface Winds: 10m Wind Vectors (
$U_{\text{wind}}, V_{\text{wind}}$ ). - Spatial Coordinates: Normalized Latitude & Longitude positional encodings.
- GLORYS Reanalysis: Subsurface temperature targets across 15 standard depths.
- ARGO Floats: In-situ CTD autonomous profiling float trajectories and profiles for independent validation.
Input Surface Satellite Observations [SST, SSS, SSH, U_curr, V_curr, U_wind, V_wind] + [Lat, Lon] (9 Channels)
↓
Spatial Coordinate & Feature Normalization
↓
Multi-Scale 2D ResNet CNN Spatial Encoder
↓
Spatial & Channel Attention Mechanism (CBAM)
↓
Ocean Latent Representation with Active Monte Carlo Dropout (p=0.15)
↓
Depth-Aware Multi-Head Prediction Decoder
↓
Subsurface Temperatures across 15 Standard Depths: [0m to 1000m]
- Real Monte Carlo Dropout with
$N=20$ forward stochastic inference passes. - Predictive Mean:
$\mu = \frac{1}{N} \sum \hat{y}_i$ - Epistemic Uncertainty:
$\sigma = \sqrt{\frac{1}{N-1}\sum (\hat{y}_i - \mu)^2}$ - Uncertainty %:
$\frac{\sigma}{|\mu|} \times 100$
OceanEmbedX/
├── backend/
│ ├── app/
│ │ ├── main.py # FastAPI application & router mounting
│ │ ├── config.py # App configuration & physical constants
│ │ ├── api/
│ │ │ ├── routes_predict.py # Predict, query, uncertainty, 3D/2D maps
│ │ │ ├── routes_datasets.py # Dataset status, upload, metadata, sample gen
│ │ │ ├── routes_validation.py # GLORYS & ARGO validation, metrics calculation
│ │ │ └── routes_train.py # PyTorch model training triggers & progress
│ │ ├── models/
│ │ │ ├── oceanembed_net.py # PyTorch OceanEmbed Architecture (CNN + Attention + Depth Head)
│ │ │ └── uncertainty.py # Monte Carlo Dropout Engine
│ │ ├── preprocessing/
│ │ │ ├── preprocess_oisst.py # NOAA OISST NetCDF parser & regridder (0.25°)
│ │ │ ├── preprocess_sss.py # SSS regridder
│ │ │ ├── preprocess_ssh.py # SSH/SLA regridder
│ │ │ ├── preprocess_currents.py # Currents regridder
│ │ │ ├── preprocess_winds.py # Winds regridder
│ │ │ ├── preprocess_glorys.py # GLORYS 15-depth target extractor
│ │ │ ├── preprocess_argo.py # ARGO profile parser & spatial-temporal matcher
│ │ │ ├── harmonize.py # Multi-source spatio-temporal alignment & normalization
│ │ │ └── sample_generator.py # Realistic synthetic NetCDF generator for BoB
│ │ ├── training/
│ │ │ ├── dataset.py # PyTorch Dataset for ocean grids
│ │ │ ├── train.py # Chronological training loop with early stopping
│ │ │ ├── evaluate.py # Scientific validation & error metrics
│ │ │ └── config.yaml # Hyperparameters & depth levels config
│ │ └── utils/
│ │ ├── metrics.py # RMSE, Bias, Pearson r, MAE, Percentage Error
│ │ └── export.py # CSV/JSON export utility
│ ├── tests/
│ │ └── test_oceanembed.py # Automated pytest suite
│ └── requirements.txt
├── frontend/
│ ├── src/
│ │ ├── components/
│ │ │ ├── Navbar.tsx # Scientific header & mode status
│ │ │ ├── Sidebar.tsx # Navigation
│ │ │ ├── QueryPanel.tsx # Lat/Lon/Depth/Date interactive picker
│ │ │ ├── PlotlyMap2D.tsx # 2D Heatmap & uncertainty viewer
│ │ │ ├── PlotlyMap3D.tsx # Interactive 3D WebGL Subsurface Map (0-1000m)
│ │ │ ├── VerticalProfileChart.tsx# 0-1000m Depth Profile (OceanEmbed vs GLORYS vs ARGO)
│ │ │ ├── ValidationCards.tsx # KPI metrics cards
│ │ │ ├── ScatterPlot.tsx # 1:1 Parity correlation scatter
│ │ │ └── DatasetUploader.tsx # NetCDF upload & simulation generator
│ │ ├── pages/
│ │ │ ├── Dashboard.tsx # Primary Showcase SIH Dashboard
│ │ │ ├── DatasetStatusPage.tsx # Data availability & file ingestion
│ │ │ ├── OceanMap2DPage.tsx # 2D horizontal depth slices
│ │ │ ├── OceanMap3DPage.tsx # Full 3D Interactive Subsurface explorer
│ │ │ ├── PredictionPage.tsx # Single-point & grid inference with MC Dropout
│ │ │ ├── ValidationPage.tsx # GLORYS Reanalysis & ARGO Float verification
│ │ │ ├── TrainingPage.tsx # Live model training monitor
│ │ │ └── AboutPage.tsx # Methodology, formulas & architecture
│ │ ├── services/api.ts # Axios API client
│ │ ├── types/index.ts # TypeScript interfaces
│ │ └── App.tsx
│ ├── package.json
│ ├── vite.config.ts
│ └── tailwind.config.js
├── README.md
└── .env.example
- Python 3.10+
- Node.js 18+ and npm
# 1. Navigate to backend
cd backend
# 2. Install dependencies
pip install -r requirements.txt
# 3. Start FastAPI server
python -m uvicorn backend.app.main:app --host 0.0.0.0 --port 8000 --reloadBackend API will run at http://127.0.0.1:8000 (Swagger docs at http://127.0.0.1:8000/docs).
# 1. Navigate to frontend
cd frontend
# 2. Install dependencies
npm install
# 3. Start development server
npm run devFrontend interface will run at http://localhost:5173.
python -m pytest backend/tests/test_oceanembed.py -v-
Select Region: Bay of Bengal (
$5^\circ\text{N}\text{ to }22^\circ\text{N}, 80^\circ\text{E}\text{ to }100^\circ\text{E}$ ). -
Select Date & Coordinate: Choose latitude, longitude, and depth level (
$0\text{--}1000\text{m}$ ). -
Load Surface Observations: Ingests SST, SSS, SSH/SLA, Surface Currents (
$U, V$ ), Surface Winds ($U, V$ ). - OceanEmbed Inference: CNN + Attention spatial embedding reconstructs depth profile across all 15 standard depths.
-
Uncertainty Quantification: MC Dropout calculates
$\pm \sigma$ and uncertainty %. -
Ground Truth Validation: Dynamically retrieves GLORYS reanalysis target and calculates RMSE, Bias, and Pearson
$r$ . -
Independent In-Situ Validation: Matches nearby ARGO profiling floats (
$<0.8^\circ$ tolerance). -
Interactive Visualizations:
- Interactive 3D Subsurface Map: Orbit, zoom, pan across depth layers with Red-Yellow-Green continuous temperature colormap.
-
Vertical Thermal Profile:
$0\text{--}1000\text{m}$ inverted depth curve with shaded$\pm 1\sigma$ MC Dropout confidence envelope. - 2D Spatial Cross-Sections: Horizontal slices of Temperature, Uncertainty, and Percentage Error.
- Data Export: Export single-point profiles or grid predictions to CSV/JSON.