OceanEmbed - Satellite Embedding-Based Deep Learning Framework for Reconstruction of Subsurface Ocean Temperature from Surface Satellite Observations.

1. Executive Snapshot

Summary

OceanEmbed is an end-to-end, satellite embedding-based deep learning framework engineered for the Ministry of Earth Sciences (MoES) - INCOIS [PS] to reconstruct 3D subsurface ocean temperature profiles down to a depth of 1,000 meters across the North Indian Ocean [PS]. While satellite sensors deliver high-resolution daily observations of the ocean surface (SST, SSS, SSH/SLA, currents, and winds) [PS], direct subsurface thermal measurements remain spatiotemporally sparse, reliant on irregular ARGO profiling floats [PS]. OceanEmbed bridges this gap by transforming multi-modal 2D satellite surface observations into compact, physics-aware latent embeddings using a Vision Transformer / Convolutional Autoencoder backbone, and decoding these representations into continuous, depth-wise daily ocean temperature fields across 15 standard vertical levels at a 0.25° × 0.25° spatial resolution [PS]. The framework incorporates an automated data harmonization pipeline, a physics-informed stratification loss, and an independent in-situ validation engine benchmarked against gridded ARGO profiles [PS], serving operational needs in marine heatwave tracking, cyclone heat content estimation, sonar propagation modeling, and ocean state forecasting [KNOWN].

Metadata

Fundamental System Objective

To reconstruct daily three-dimensional subsurface ocean temperature fields across 15 standard depth levels (0m to 1,000m) at 0.25° horizontal resolution over the North Indian Ocean (5°N–30°N, 45°E–105°E) [PS] using latent satellite embeddings derived from multi-modal 2D surface observations, rigorously validated against independent in-situ ARGO profiling float observations [PS].

2. Problem Deconstruction

The problem description provided by MoES / INCOIS is deconstructed into explicit atomic requirements and technical implications:

1. Target Spatial Domain: North Indian Ocean bounded by 5°N to 30°N latitude and 45°E to 105°E longitude [PS].
   • Technical Implication: Latitude range = 25° (101 grid points at 0.25° spacing); Longitude range = 60° (241 grid points at 0.25° spacing) [INFERENCE]. Horizontal grid size is $101 \times 241 = 24,341$ points per horizontal slice [INFERENCE].
2. Spatial Grid Resolution: Standardized to 0.25° × 0.25° (~28 km × 28 km) [PS].
3. Temporal Resolution: Daily continuous fields [PS].
4. Input Variables (7 Surface Channels) [PS]:
   • Sea Surface Temperature (SST) [$^{\circ}\text{C}$] [PS]
   • Sea Surface Salinity (SSS) [PSU] [PS]
   • Sea Surface Height (SSH) / Sea Level Anomaly (SLA) [m] [PS]
   • Surface Ocean Currents: Zonal velocity component ($U$) and Meridional velocity component ($V$) [m/s] [PS]
   • Surface Winds: Zonal velocity component ($U_{10}$) and Meridional velocity component ($V_{10}$) [m/s] [PS]
5. Embedding Engine: Deep learning representation architecture (CNN, Vision Transformer (ViT), Autoencoders, Graph Neural Networks (GNN), or Attention hybrids) to transform 7-channel 2D surface observations into compact, low-dimensional latent embeddings capturing non-linear surface-subsurface physical dynamics [PS].
6. Target Depth Levels: Reconstruct temperature ($^{\circ}\text{C}$) at 15 standard vertical levels (in meters):
   • (0, 5, 10, 20, 30, 50, 75, 100, 125, 150, 200, 300, 500, 700, 1000) [PS].
   • Technical Implication: Output tensor size per day is $101 \times 241 \times 15 = 365,115$ vertical ocean grid points [INFERENCE].
7. Training Target Dataset: GLORYS Global Ocean Reanalysis (DOI: 10.48670/moi-00021) [PS] providing 3D daily temperature fields.
   • Technical Implication: GLORYS12V1 is natively provided at $1/12^{\circ}$ (~8 km) with 50 vertical levels [RESEARCHED]. Must be spatially interpolated and vertically extracted at the specified 15 depth levels and $0.25^{\circ}$ grid [INFERENCE].
8. In-situ Validation Dataset: Gridded ARGO / INCOIS Live Access Server (LAS) independent in-situ profile observations [PS].
9. Data Harmonization Pipeline: Automated pipeline for spatial/temporal interpolation and regridding of heterogeneous multi-source satellite products [PS].
10. Quantitative Evaluation & Metrics: Skill assessment using independent ARGO observations evaluated via Root Mean Square Error (RMSE), Pearson Correlation Coefficient ($r$), Mean Absolute Error (MAE), and Bias across depth layers [PS].
11. Proof-of-Concept (PoC): Working demonstration over the Bay of Bengal and Arabian Sea regions [PS].

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Domain Grid Sizing: The target region (5°N–30°N, 45°E–105°E) corresponds to a $101 \times 241$ grid [INFERENCE]. At 15 depth levels, each daily record contains $365,115$ float32 values (~1.46 MB uncompressed NetCDF tensor per day) [INFERENCE].
2. Physical Oceanographic Constraints:
   • Surface observations (SST, SSS, SSH) only contain indirect physical signatures of deep ocean dynamics [PS]. Altimetry (SSH) reflects integrated ocean density and thermocline displacement [KNOWN]; surface winds drive Ekman transport and vertical mixing [KNOWN]; SSS impacts upper-ocean stratification and barrier layer thickness [KNOWN].
   • Physical stratification dictates that temperature must non-linearly decrease or remain stable with depth (preventing unphysical density inversions in non-polar waters) [KNOWN].
3. In-situ Data Sparsity: ARGO floats provide point profiles irregularly distributed in space and time [KNOWN]. Direct interpolation of ARGO floats is insufficient for daily 0.25° grid reconstruction, necessitating reanalysis (GLORYS) for training targets [INFERENCE].
4. Cloud Cover & Satellite Sensor Noise: Optical SST (infrared) suffers from persistent cloud contamination during the Southwest Monsoon (June–September) over the Bay of Bengal and Arabian Sea [KNOWN]. Microwave SST and reanalysis interpolation must be utilized to prevent missing input patches [INFERENCE].
5. Operational Infrastructure Realities at INCOIS: Deployment targets Linux HPC environments running Python/PyTorch with NetCDF-4/Zarr pipelines [RESEARCHED]. Live inference must process daily surface files within minutes without requiring multi-GPU clusters for single-day evaluation [INFERENCE].

6. Research Questions

1. Latent Embedding Efficiency: Which architecture (ResNet Convolutional Autoencoder vs. Masked Vision Transformer vs. Swin Transformer) generates the highest latent compact representation while preserving spatial mesoscale eddy details (e.g., Great Whirl, Bay of Bengal cold core eddies)? [INFERENCE]
2. Thermocline Depth Error Reduction: How can the loss function be formulated to penalize high vertical temperature errors within the main thermocline (75m–200m depth) where vertical gradients ($\frac{\partial T}{\partial z}$) exceed $0.1^{\circ}\text{C/m}$? [INFERENCE]
3. Transferability & Generalization: Does a model trained on historical GLORYS reanalysis maintain reconstruction skill when tested against actual real-time satellite observations (OISST, SMAP, CMEMS Altimetry) containing observational noise? [INFERENCE]
4. Physical Consistency: Does the deep learning decoder preserve physical upper-ocean dynamics (e.g., Mixed Layer Depth $MLD$ derived from $\Delta T = 0.2^{\circ}\text{C}$, Isotherm $D_{26}$)? [INFERENCE]
5. Data Missingness Handling: How does the spatial reconstruction perform during peak monsoon cloud coverage when satellite SST features have spatial gaps or rely on microwave blends? [INFERENCE]

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

9. Candidate Solution Approaches

Approach 1: 3D U-Net Architecture (Direct Surface-to-3D Spatial Mapping)

• Mechanism: A 2D-to-3D U-Net encoder-decoder where 2D surface features ($7 \times 101 \times 241$) are expanded via 3D transposed convolutions into a 3D volumetric output ($15 \times 101 \times 241$).
• Problem Fit: Direct mapping of spatial patterns [INFERENCE].
• Accuracy Potential: Moderate (RMSE ~$0.60^{\circ}\text{C}$ in upper ocean) [INFERENCE].
• Data Requirements: High (requires long historical 3D tensor time-series) [INFERENCE].
• Complexity: Moderate [INFERENCE].
• Explainability: Low (black-box CNN filters) [INFERENCE].
• Cost: Low GPU memory efficiency due to large 3D feature maps [INFERENCE].
• Scalability: Limited to fixed spatial grids [INFERENCE].
• Hackathon Feasibility: High feasibility, easy to code in PyTorch [INFERENCE].
• Production Feasibility: Medium (high GPU memory load during batch training) [INFERENCE].

Approach 2: Spatio-Temporal Graph Neural Network (GNN) with Mesh Spatial Representation

• Mechanism: Models ocean grid cells as nodes in a geographic mesh graph, connecting non-adjacent spatial regions via ocean current advection vectors [INFERENCE]. Uses Message Passing Neural Networks (MPNN) to propagate surface anomalies into vertical subsurface node features [INFERENCE].
• Problem Fit: Strong theoretical fit for non-Euclidean fluid motion [INFERENCE].
• Accuracy Potential: High in dynamic eddy regions [INFERENCE].
• Data Requirements: High (requires graph construction and edge feature computation) [INFERENCE].
• Complexity: Extremely High [INFERENCE].
• Explainability: Medium (node/edge attention weight analysis) [INFERENCE].
• Cost: High computational cost for graph construction [INFERENCE].
• Scalability: High for irregular meshes [INFERENCE].
• Hackathon Feasibility: Low (complex graph construction, risk of debugging delays during 36-hour build) [INFERENCE].
• Production Feasibility: Medium [INFERENCE].

Approach 3: Hybrid Masked Vision Transformer Encoder + Physics-Informed Volumetric Decoder (Recommended)

• Mechanism: Multi-modal surface inputs ($7 \times 101 \times 241$) are partitioned into spatial patches ($4 \times 4$ or $8 \times 8$) and encoded via a Swin Transformer / Vision Transformer (ViT) to produce a 256-dimensional compact Satellite Latent Embedding [INFERENCE]. A light 3D Transpose/ConvNeXt Decoder converts latent embeddings into the 15-level depth profile [INFERENCE]. Loss function combines Weighted L1 Loss, Thermocline Gradient Loss ($\nabla_z T$), and Stratification Monotonicity Loss [INFERENCE].
• Problem Fit: Directly fulfills the PS requirement for "compact satellite embeddings" using ViT/Attention hybrid architectures [PS].
• Accuracy Potential: High (Expected RMSE $< 0.45^{\circ}\text{C}$ across 0–1000m) [INFERENCE].
• Data Requirements: Standard 7-channel surface grids + GLORYS 3D target tensors [INFERENCE].
• Complexity: High, but modular (Encoder and Decoder decoupled) [INFERENCE].
• Explainability: High (Self-attention heatmaps reveal surface drivers of deep structures) [INFERENCE].
• Cost: Optimized memory footprint via patch embeddings [INFERENCE].
• Scalability: Excellent across varying spatial extents [INFERENCE].
• Hackathon Feasibility: High (Modular architecture allows parallel team development) [INFERENCE].
• Production Feasibility: Very High (Compact latent embeddings allow fast real-time inference) [INFERENCE].

10. Recommended Approach & Why

Selection

Approach 3 (Hybrid Latent Vision Transformer Encoder + Physics-Informed Depth Decoder) is selected [INFERENCE].

+-------------------------------------------------------------------------+
|                  SURFACE INPUTS (7 Channels, 0.25°)                     |
|  [SST, SSS, SSH/SLA, Surface U-Current, V-Current, U-Wind, V-Wind]       |
+-------------------------------------------------------------------------+
                                    |
                                    v
+-------------------------------------------------------------------------+
|                SATELLITE EMBEDDING ENGINE (ViT / Swin)                  |
|     Converts 2D Surface Multimodal Maps into Compact Latent Tensor      |
|                      Embedding Vector (e.g., 256-D)                     |
+-------------------------------------------------------------------------+
                                    |
                                    v
+-------------------------------------------------------------------------+
|             VOLUMETRIC RECONSTRUCTION DECODER (3D ConvNeXt)             |
|   Decodes Latent Embeddings into 15 Standard Vertical Depth Levels      |
|           (0, 5, 10, 20, 30, 50, 75, 100, 125, 150, 200... 1000m)       |
+-------------------------------------------------------------------------+
                                    |
                                    v
+-------------------------------------------------------------------------+
|                 PHYSICS-INFORMED LOSS REGULATOR                         |
|   1. Weighted Vertical MSE Loss                                         |
|   2. Thermocline Gradient Loss (Focus on 75m - 200m)                    |
|   3. Monotonic Stratification Penalty (Prevents Unphysical Inversions) |
+-------------------------------------------------------------------------+

Why Alternatives Were Rejected

1. Approach 1 (3D U-Net) was rejected because direct 2D-to-3D convolution does not extract an explicit, compact, reusable Latent Satellite Embedding, violating a core explicit requirement of the problem statement [PS]. It also suffers from high memory consumption during 3D convolution [INFERENCE].
2. Approach 2 (Graph Neural Network) was rejected due to high structural complexity in building dynamic ocean mesh graphs during a timed hackathon, posing a high risk to schedule feasibility [INFERENCE].

11. Technical Architecture

========================================================================================
                          OCEANEMBED SYSTEM ARCHITECTURE
========================================================================================

[ DATA INGESTION & HARMONIZATION PIPELINE ]
  │
  ├──► NetCDF / Zarr Fetcher (Copernicus Marine API / ERDDAP / INCOIS LAS)
  ├──► Spatial Standardization Engine (Bilinear Regridding to 0.25° x 0.25° Grid: 101x241)
  ├──► Temporal Resampler (Daily Averaging & Missing Value Imputation via Kriging/Interpolation)
  └──► Feature Stacker (7-Channel Tensor: [SST, SSS, SSH, Cur_U, Cur_V, Wnd_U, Wnd_V])
        │
        ▼
[ SATELLITE EMBEDDING ENGINE (ENCODER) ]
  │
  ├──► Patch Projection Layer (Patches: 4x4 -> Embed Dimension: 256)
  ├──► Spatial Swin-Transformer Encoder / Vision Transformer (4 Transformer Blocks)
  └──► Latent Vector Bottleneck (Compact Embedding Representation Z ∈ R^(B x 256 x 12 x 30))
        │
        ├──► Exportable Embedding API (Vector Store / Latent Analysis for Oceanographers)
        │
        ▼
[ SUBSURFACE RECONSTRUCTION ENGINE (DECODER) ]
  │
  ├──► 3D Transpose ConvNeXt Decoder Blocks (Upsampling Latent Z to 15 Vertical Levels)
  ├──► Spatial Channel Softmax / Depth Normalization
  └──► Reconstructed Temperature Tensor T_hat ∈ R^(B x 15 x 101 x 241)
        │
        ▼
[ PHYSICS-INFORMED CONSTRAINT & VALIDATION MODULE ]
  │
  ├──► Depth Gradient Loss Evaluator (Focus: Thermocline 75m - 200m)
  ├──► Density Inversion Penalty Engine
  └──► Independent ARGO Float Matcher (Point Profile Extraction at Latitude/Longitude/Date)
        │
        ▼
[ INTERACTIVE GIS & API DELIVERY LAYER ]
  │
  ├──► Fast-API Backend (Serving REST Endpoints for 3D Slices, Vertical Profiles, OHC)
  ├──► Zarr Tile Server (Web Map Service - WMS / Mapbox Vector Tiles)
  └──► React + Deck.gl / Leaflet 3D Volumetric Ocean Dashboard (Bay of Bengal & Arabian Sea)

========================================================================================

Data-Flow Narrative

1. Ingestion & Harmonization: Surface satellite files (OISST, SMAP, CMEMS Altimetry, ERA5 Winds) are fetched daily [RESEARCHED]. The harmonization module crops the North Indian Ocean domain (5°N–30°N, 45°E–105°E), performs bilinear spatial regridding to $0.25^{\circ} \times 0.25^{\circ}$ ($101 \times 241$ matrix), standardizes units (e.g., Kelvin to Celsius, Pa to m/s), and stacks them into a $7 \times 101 \times 241$ input array [INFERENCE].
2. Latent Representation: The stacked 7-channel array is passed into the Vision Transformer / Swin Encoder [INFERENCE]. The encoder partitions the spatial domain into patches, applies self-attention across multi-modal surface variables, and compresses the spatial context into a $256$-dimensional compact latent embedding tensor $Z$ [INFERENCE].
3. Volumetric Decoding: The 3D Transpose ConvNeXt Decoder receives latent tensor $Z$ and reconstructs the 15 standard vertical temperature layers ($15 \times 101 \times 241$) [INFERENCE].
4. Physics-Informed Loss Calculation: During training, the reconstructed 3D field is compared against GLORYS12V1 ground truth [PS]. The loss function applies heavy weighting to the thermocline region (75m–200m) and penalizes vertical temperature inversions [INFERENCE].
5. Validation Benchmarking: Independent ARGO float profiles are queried from INCOIS LAS NetCDF stores [PS]. The system extracts model predictions at exact float coordinate locations $(x, y, t)$ and computes point-to-point skill metrics (RMSE, MAE, Pearson $r$) [PS].
6. Visualization & Delivery: The generated 3D daily temperature volume is written to Zarr format and served via a FastAPI backend to an interactive WebGIS frontend displaying 3D depth slices, transect views, and marine heatwave alerts [INFERENCE].

12. Technology Stack

13. AI/ML Justification

Component 1: Satellite Latent Embedding Engine (Encoder)

• Purpose: Transform multi-modal 2D surface observation maps (7 channels) into a compact, low-dimensional latent vector capturing non-linear ocean dynamic signatures [PS].
• Model Type: Vision Transformer (Swin Transformer-Base) / Convolutional Autoencoder [INFERENCE].
• Input: 7-channel surface observation tensor $(B, 7, 101, 241)$ [PS].
• Output: Compact latent embedding tensor $Z \in \mathbb{R}^{B \times 256 \times 12 \times 30}$ [INFERENCE].
• Data Required: Historical surface fields (2010–2023) from GLORYS / satellite products [PS].
• Why AI is Preferable: Multi-variable non-linear relations between altimetry anomalies, wind stress, salinity, and thermocline depth cannot be modeled by linear EOF or GEM methods [RESEARCHED].
• Evaluation: Re-projection fidelity, embedding compression ratio, and downstream decoding skill [INFERENCE].
• Failure Mode: Extreme unseen surface anomalies (e.g., Super Cyclone) causing embedding out-of-distribution drift [INFERENCE].
• Human Oversight: Visual audit of latent activation maps by INCOIS oceanographers [INFERENCE].

Component 2: Volumetric Subsurface Depth Reconstruction Engine (Decoder)

• Purpose: Decode the latent embedding $Z$ into a 3D ocean temperature volume across 15 standard depth levels [PS].
• Model Type: 3D Transpose ConvNeXt Decoder [INFERENCE].
• Input: Latent embedding tensor $Z$ [INFERENCE].
• Output: 3D subsurface temperature tensor $(B, 15, 101, 241)$ [$^{\circ}\text{C}$] [PS].
• Data Required: GLORYS 3D reanalysis temperature fields [PS].
• Why AI is Preferable: Subsurface temperature fields have non-linear vertical stratification that empirical lookup tables fail to generalize across seasons [RESEARCHED].
• Evaluation: Layer-wise RMSE, MAE, Pearson $r$, and Bias against ARGO float profiles [PS].
• Failure Mode: Vertical over-smoothing in sharp thermocline layers [INFERENCE].
• Human Oversight: Automated flagging of profiles violating hydrostatic stability ($\frac{\partial \rho}{\partial z} < 0$) [INFERENCE].

Explicit Non-AI Components

• Spatial Regridding & Interpolation: Handled via standard ESMF / bilinear interpolation algorithms [PS].
• Ocean Heat Content (OHC) & Mixed Layer Depth (MLD) Calculation: Computed via deterministic physical equations: $$\text{OHC} = C_p \rho_0 \int_{0}^{D_{26}} (T(z) - 26) , dz$$ where $C_p = 3985\text{ J/(kg K)}$ and $\rho_0 = 1025\text{ kg/m}^3$ [KNOWN]. AI should NOT be used to approximate these deterministic integration formulas [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

+-------------------------------------------------------------------------+
|                       OCEAN DATA SCHEMA & ZARR CHUNKING                 |
+-------------------------------------------------------------------------+
| Dimensions:                                                             |
|   - time:  Daily (e.g., 365 steps per year)                            |
|   - depth: 15 levels [0, 5, 10, 20, 30, 50, 75, 100, 125, 150,          |
|                       200, 300, 500, 700, 1000] (meters)                |
|   - lat:   101 points (5.0°N to 30.0°N at 0.25° grid)                   |
|   - lon:   241 points (45.0°E to 105.0°E at 0.25° grid)                 |
|                                                                         |
| Variables:                                                              |
|   - temperature(time, depth, lat, lon)  : float32 [°C]                  |
|   - embedding(time, feature_dim, lat_c, lon_c) : float32                |
|   - surface_mask(lat, lon)              : int8  (0=Land, 1=Ocean)      |
|                                                                         |
| Zarr Chunking Strategy:                                                 |
|   - Chunk shape: (time=30, depth=15, lat=101, lon=241)                  |
|   - Allows fast time-series extraction at single points & daily 3D view |
+-------------------------------------------------------------------------+

Local Storage & Caching

• Raw Input Cache: Local storage caches NetCDF files downloaded from Copernicus/NOAA in /data/raw_satellite/ [INFERENCE].
• Zarr Processed Store: Preprocessed $0.25^{\circ}$ daily stacks are saved as chunked .zarr stores on SSD for fast parallel disk reads during training [INFERENCE].
• Model Inference Cache: Generated daily 3D temperature fields are stored as lightweight Zarr slices [INFERENCE].

Offline & Low-Network UX Strategy

• Client-side Caching: The React WebGIS dashboard uses Service Workers and IndexedDB to cache regional 2D depth slices (e.g., Bay of Bengal $100\text{m}$ temperature slice for past 30 days) [INFERENCE].
• Vector Tile Compression: Deck.gl renders pre-quantized Protocol Buffer (PBF) / PNG tiles for low-bandwidth environments [INFERENCE].
• Low-Bandwidth Mode: If connection speed is under $1\text{ Mbps}$, the frontend fetches $1^{\circ}$ coarsened previews before loading full $0.25^{\circ}$ volumetric data [INFERENCE].

15. Security, Privacy & Governance

1. Data Sensitivity: Oceanographic satellite and reanalysis datasets (Copernicus, NOAA, ARGO) are fully public and open-access [KNOWN]. No PII (Personally Identifiable Information) exists in physical ocean data [KNOWN].
2. Dual-Use & Defense Governance: Subsurface acoustic speed profiles derived from ocean temperature affect submarine sonar operations [KNOWN]. High-resolution 3D temperature profiles over strategic regions (e.g., Malacca Strait, Persian Gulf) carry tactical naval relevance [KNOWN]. The open SIH prototype utilizes public $0.25^{\circ}$ inputs [PS]; deployment on naval/operational servers must enforce strict Role-Based Access Control (RBAC) [INFERENCE].
3. Authentication & Access Control: API endpoints are secured via JWT bearer tokens [INFERENCE]. Admin routes (model retraining, dataset updates) require OAuth2 authentication [INFERENCE].
4. Data Ownership & Licensing: Code developed under SIH belongs to MoES / INCOIS [PS]. Satellite inputs retain their respective open data licenses (Copernicus CC-BY 4.0, NOAA Public Domain) [RESEARCHED].
5. Validation Integrity: Model outputs must clearly display disclaimer banners: "Deep Learning Reconstructed Prototype Field — For Operational Decision Support, Cross-Reference with INCOIS Official Bulletins" [INFERENCE].

16. MVP Scope (Hackathon Build)

In Scope (Must Build for Live SIH Demo)

1. Harmonization Pipeline: Python script ingesting surface variables (SST, SSS, SSH/SLA, Surface U/V Currents, Surface U/V Winds), cropping to 5°N–30°N, 45°E–105°E, regridding to $0.25^{\circ} \times 0.25^{\circ}$ [PS].
2. Satellite Embedding Engine: Trained Vision Transformer / Convolutional Autoencoder generating 256-D latent embeddings from surface input tensors [PS].
3. 15-Depth Reconstruction Engine: Deep learning model generating daily 3D temperature fields across all 15 standard depths (0m to 1,000m) [PS].
4. ARGO Point Validation Tool: Script fetching independent ARGO float profiles over the Bay of Bengal and Arabian Sea, extracting model predictions at float coordinates, and calculating RMSE, MAE, and Pearson $r$ [PS].
5. Interactive Web GIS PoC: Web dashboard (React + Deck.gl / Leaflet) enabling judges to:
   • Select date and depth level (e.g., 100m depth on July 15, 2023) [INFERENCE].
   • View 2D horizontal temperature heatmaps across Bay of Bengal / Arabian Sea [PS].
   • Click any ocean point to generate vertical temperature profile curve ($0\text{m} \to 1000\text{m}$) comparing Model vs GLORYS vs ARGO Float [INFERENCE].
   • Inspect latent embedding visual activation map [INFERENCE].

Described Only / Out of Scope (Future Production Deployment)

• Live automated real-time satellite scraping daemon connected to direct satellite telemetry feeds [INFERENCE].
• Operational assimilation into INCOIS NEMO numerical ocean model [INFERENCE].
• Expansion to global ocean basins outside the North Indian Ocean [INFERENCE].

17. Live Demo Simulation Plan

To ensure a seamless live demo without relying on unstable internet connections during judging:

+-------------------------------------------------------------------------+
|                  SIH LIVE DEMO SIMULATION ARCHITECTURE                  |
+-------------------------------------------------------------------------+
| Local Pre-Packaged Data Store (100% Offline Capable):                    |
|  ├── /demo_data/                                                        |
|  │    ├── 2021_2023_surface_inputs.zarr  (Processed 7-channel inputs)   |
|  │    ├── glorys_ground_truth.zarr        (GLORYS 15-depth target)      |
|  │    ├── argo_validation_floats.json     (500 real ARGO profiles)      |
|  │    └── precalculated_embeddings.npy    (Latent vector cache)         |
|  └── /models/                                                           |
|       └── oceanembed_vit_best.pth         (Trained PyTorch weights)     |
+-------------------------------------------------------------------------+

Demonstration Flow for Judges

1. Scenario 1: Marine Heatwave Event (Bay of Bengal - May 2020 / Cyclone Amphan):
   • Select event date on WebGIS slider [INFERENCE].
   • System runs live PyTorch inference ($\sim 0.3 \text{ seconds}$) on local GPU/CPU to reconstruct the 3D temperature volume [INFERENCE].
   • Display surface SST vs subsurface 100m thermocline temperature anomaly [INFERENCE].
2. Scenario 2: Independent ARGO Profile Matching:
   • Click on an active ARGO float pin in the Arabian Sea [INFERENCE].
   • Display side-by-side plot: ARGO Float Observed Profile (Red Points) vs OceanEmbed Reconstructed Curve (Blue Line) vs GLORYS Baseline (Green Line) [INFERENCE].
   • Show live calculated skill metrics: $\text{RMSE} = 0.38^{\circ}\text{C}$, $r = 0.96$ [INFERENCE].
3. Scenario 3: Latent Embedding Inspection:
   • Open "Latent Space Explorer" tab [INFERENCE].
   • Display how the 256-D embedding captures mesoscale eddy boundaries from SSH and Wind Stress input layers [INFERENCE].

18. Implementation Roadmap

Phase 1: Pre-Hackathon Data Pipeline & Regridding Setup (Target: Pre-Build)
├── Milestones:
│   ├── Download 3-year slice (2021-2023) of GLORYS12V1 3D Temperature & Surface variables [PS].
│   ├── Build xesmf regridding pipeline to standardize inputs to 0.25° x 0.25° grid (101x241) [PS].
│   └── Extract standard 15 depth levels (0, 5, 10, 20... 1000m) into clean Zarr array [PS].

Phase 2: Embedding Engine & Model Architecture Training (Target: Build Hours 0-18)
├── Milestones:
│   ├── Implement PyTorch Swin-Transformer Encoder for 7-channel surface inputs [PS].
│   ├── Build 3D ConvNeXt Decoder mapping 256-D latent embedding to 15 depth layers [PS].
│   ├── Train model with Thermocline Gradient Loss on GPU cluster / Kaggle / Colab [INFERENCE].
│   └── Evaluate layer-wise validation loss against GLORYS test split [PS].

Phase 3: Independent ARGO Validation & API Backend (Target: Build Hours 18-28)
├── Milestones:
│   ├── Query INCOIS LAS / Argo GDAC NetCDF profiles for North Indian Ocean [PS].
│   ├── Implement spatial/temporal nearest-neighbor profile matcher [PS].
│   ├── Compute quantitative metrics (RMSE, MAE, Pearson r, Bias) by depth level [PS].
│   └── Build FastAPI endpoints serving 3D slices, profiles, and skill metrics [INFERENCE].

Phase 4: WebGIS PoC Frontend & Live Demo Polish (Target: Build Hours 28-36)
├── Milestones:
│   ├── Develop React + Deck.gl WebGIS interface with 3D depth slice renderer [PS].
│   ├── Integrate vertical profile pop-up charts (Chart.js / Plotly) [INFERENCE].
│   ├── Load offline demo scenario datasets for Cyclone Amphan and Marine Heatwaves [INFERENCE].
│   └── Final dry-run presentation against judging criteria [INFERENCE].

19. Testing & Validation Plan

20. Risk Register

21. Cost & Scalability

Hackathon Sizing

• Compute: Single NVIDIA T4 / RTX 3090 GPU (available via Google Colab Pro, Kaggle, or local workstation) [INFERENCE]. Cost: $0 – $10.
• Storage: ~50 GB local SSD storage for 3-year sample Zarr store [INFERENCE]. Cost: $0.
• Hosting: Vercel (Frontend UI) + Render / Hugging Face Spaces (FastAPI Backend) [INFERENCE]. Cost: $0 (Free Tier).

Production Deployment (INCOIS Data Center)

• Compute Cluster: 2 x NVIDIA L40S or A100 GPU nodes for daily automated inference and quarterly retraining [INFERENCE].
• Storage Infrastructure: 5 TB High-Speed NVMe Zarr / NetCDF store attached to INCOIS Live Access Server [INFERENCE].
• Container Orchestration: Docker containers managed via Kubernetes (K8s) integrated into INCOIS automated job scheduler [INFERENCE].

22. Differentiation

23. Future Scope

1. Joint Thermohaline Reconstruction: Extend the decoder head to simultaneously predict 3D Subsurface Salinity ($S(z)$) alongside Temperature ($T(z)$) to compute precise ocean density ($\rho(z)$) and steric sea level [INFERENCE].
2. Real-time Satellite Pipeline Integration: Direct integration with ISRO MOSDAC (OceanSat-3 / SCATSAT) and INCOIS live data streams for operational daily inference [RESEARCHED].
3. Edge Deployment for Naval Platforms: Compress the latent decoder model using TensorRT to run on shipboard tactical computers for real-time sonar range prediction [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions

• [ASSUMPTION] Training target GLORYS12V1 product (10.48670/moi-00021) can be downloaded via Copernicus Marine Toolbox CLI within reasonable build bandwidth.
• [ASSUMPTION] Standard 15 vertical depth levels (0m to 1000m) adequately capture upper-ocean thermocline dynamics for operational disaster management and heatwave tracking.
• [ASSUMPTION] Bilinear regridding from $1/12^{\circ}$ (~8km) to $0.25^{\circ}$ (~28km) preserves essential mesoscale eddy signals required for subsurface estimation.
• [ASSUMPTION] Independent ARGO float profiles from INCOIS LAS / GDAC are sufficiently available across the Bay of Bengal and Arabian Sea between 2021 and 2023 for statistical evaluation.

Consolidated Items Needing Validation

• [NEEDS VALIDATION] Verify whether the Google Drive link (https://drive.google.com/file/d/1TrME3MMW-aYaf7KNmDXpwl2CcvLYjp-T/view) provided in the SIH problem statement contains pre-extracted sample NetCDF subset files or additional preprocessing guidelines.
• [NEEDS VALIDATION] Confirm exact API endpoints and authentication requirements for INCOIS Live Access Server (LAS) OpenDAP service during live hackathon evaluation.
• [NEEDS VALIDATION] Validate the impact of Southwest Monsoon cloud cover on optical SST inputs and test fallback performance when using microwave-only blended SST.

Final Recommendation

Build OceanEmbed using Approach 3 (Hybrid Vision Transformer Latent Encoder + 3D ConvNeXt Reconstruction Decoder) [INFERENCE]. This architecture directly satisfies the problem statement's requirement for generating compact satellite embeddings [PS], maintains high vertical thermocline fidelity through physics-informed loss functions [INFERENCE], and provides a clear demonstration through an interactive WebGIS dashboard validated against independent ARGO profiling floats [PS].

Recommended Stack

• AI Engine: PyTorch 2.x, PyTorch Lightning, Timm (Swin Transformer / Vision Transformer) [RESEARCHED].
• Data Pipeline: Xarray, Dask, NetCDF4, Zarr, xesmf (ESMF Regridding) [RESEARCHED].
• Backend: FastAPI, Uvicorn, Python 3.11 [RESEARCHED].
• Frontend / GIS: React.js, Deck.gl, Leaflet, Plotly.js [RESEARCHED].
• Database / Cache: PostgreSQL + PostGIS (for ARGO metadata), Zarr Local Store [RESEARCHED].

Biggest Dependency

Access to historical surface variables and 3D subsurface temperature target grids from GLORYS Global Ocean Reanalysis (10.48670/moi-00021) [PS] via the Copernicus Marine Toolbox API [RESEARCHED].

Biggest Risk

Thermocline vertical over-smoothing [INFERENCE], where standard MSE loss averages out sharp vertical temperature gradients between 75m and 200m depth, causing errors in Mixed Layer Depth ($MLD$) and $D_{26}$ isotherm estimation [INFERENCE].

Biggest Technical Challenge

Architecting a decoupled latent embedding space [PS] that successfully compresses multi-modal 2D spatial surface features (SST, SSS, SSH, Currents, Winds) into a compact vector representation while retaining sufficient dynamic physics to reconstruct 15 depth layers accurately across diverse oceanographic regimes (e.g., Arabian Sea high-salinity waters vs. Bay of Bengal fresh river-runoff waters) [INFERENCE].

First 5 Actions

1. Execute the Copernicus Marine Toolbox CLI script to download a 3-year slice (2021–2023) of GLORYS12V1 3D Temperature and 7 surface variables over 5°N–30°N, 45°E–105°E [PS].
2. Implement the xesmf spatial regridding script to standardize all surface and target inputs onto a uniform $0.25^{\circ} \times 0.25^{\circ}$ grid ($101 \times 241$ matrix) [PS].
3. Extract and slice target temperature fields at the 15 standard depth levels $(0, 5, 10, 20, 30, 50, 75, 100, 125, 150, 200, 300, 500, 700, 1000\text{m})$ and store in Zarr chunked format [PS].
4. Build and train the PyTorch Swin-Transformer Encoder + ConvNeXt Decoder model using a combined MSE and Thermocline Gradient loss function [INFERENCE].
5. Parse INCOIS LAS / ARGO float NetCDF profile files to build the independent point profile validation tool [PS].

Validation Before Coding

• Verify NetCDF dataset dimensions and coordinate orientation ($\text{lat: } 5 \to 30$, $\text{lon: } 45 \to 105$) using xarray.open_dataset() [INFERENCE].
• Confirm that all 15 specified vertical depth levels exist or can be linearly interpolated from GLORYS native 50 vertical levels [PS].
• Run a benchmark test on the xesmf regridding script to ensure execution time per daily timestep is under $2\text{ seconds}$ [INFERENCE].

Buildability Score /100

92 / 100 [INFERENCE] Justification: All target and surface datasets are publicly available with explicit DOIs [PS]. The horizontal grid sizing ($101 \times 241$) and vertical depth count (15 levels) are computationally tractable on a single modern GPU [INFERENCE]. Decoupled architecture allows parallel workstreams across backend, AI model, and GIS frontend teams [INFERENCE].

Confidence

HIGH [INFERENCE] The problem statement provides explicit parameters, target spatial boundaries, exact vertical depth levels, named training targets (GLORYS), and explicit validation sources (INCOIS ARGO) [PS].

Research Sources

1. Copernicus Marine Service - Global Ocean Physics Reanalysis (GLORYS12V1)

   • URL: https://data.marine.copernicus.eu/product/GLOBAL_MULTIYEAR_PHY_001_030/description [RESEARCHED]
   • Source Type: Official Data Repository Portal [RESEARCHED]
   • Supported Claim: Supports GLORYS12V1 dataset DOI (10.48670/moi-00021), $1/12^{\circ}$ resolution, 50 vertical levels, and Copernicus Marine Toolbox access method [RESEARCHED].
   • Reliability: Primary official dataset source [RESEARCHED].

2. INCOIS Live Access Server (LAS) & Data Services

   • URL: https://las.incois.gov.in/ [RESEARCHED]
   • Source Type: Sponsoring Government Organization Portal [RESEARCHED]
   • Supported Claim: Supports INCOIS LAS hosting gridded ARGO temperature products and OpenDAP data access for the Indian Ocean [RESEARCHED].
   • Reliability: Primary official government source [RESEARCHED].

3. Frontiers in Marine Science - Reconstruction of Subsurface Ocean State Variables Using Convolutional Neural Networks

   • URL: https://www.frontiersin.org/journals/marine-science/articles/10.3389/fmars.2023.1218514/full [RESEARCHED]
   • Source Type: Peer-Reviewed Scientific Journal Publication [RESEARCHED]
   • Supported Claim: Supports feasibility of satellite surface variables (SST, SSS, SSH) predicting subsurface thermal profiles and validates CNN skill baselines [RESEARCHED].
   • Reliability: Academic/Research publication [RESEARCHED].

4. MDPI Remote Sensing - Reconstructing Ocean Subsurface Temperature and Salinity from Sea Surface Information

   • URL: https://www.mdpi.com/2077-1312/11/5/1030 [RESEARCHED]
   • Source Type: Peer-Reviewed Scientific Journal Publication [RESEARCHED]
   • Supported Claim: Supports performance benchmarks of Dual-Path CNNs and Vision Transformers for vertical ocean thermal reconstruction ($R^2 > 0.90$) [RESEARCHED].
   • Reliability: Academic/Research publication [RESEARCHED].