nasa-cisto-data-science-group/downstream-ocean-color-subset

Dataset Card for SatVision Ocean Color Chips

Dataset Summary

The SatVision Ocean Color Chips dataset is a curated collection of multi-band satellite image chips for training and evaluating models on ocean color retrieval tasks. It is tailored for the SatVision-TOA foundation model framework and includes data prepared for both fine-tuning and end-to-end model training.

Two primary subsets are provided:

• Fine-tuned Chips (FT): Preprocessed for fine-tuning SatVision-TOA encoders with custom decoders.
• End-to-End Chips (E2E): Chips for training U-Net or similar architectures for direct retrieval.

---

Supported Tasks and Benchmarks

• Multispectral Regression: Retrieval of biogeophysical ocean parameters from 12–14 band inputs.
• Pretraining & Fine-tuning: Designed for adapting SatVision-TOA encoders to downstream tasks.

Reported metrics:

• Loss (MSE)
• MAE
• R²
• RMSE
• SSIM
• PSNR

---

Dataset Structure

• Input: Multi-band image chips (12–14 channels) in .npy or .tif format.
• Target: Single-channel continuous values representing ocean color properties.

Directories:

• training_chips_ft – Training chips for fine-tuning.
• val_chips_ft – Validation chips for fine-tuning.

Size:
~144 training samples and 37 validation samples per subset.

---

Data Fields

• image (np.ndarray): Multi-band array of shape (C, H, W) where C ∈ {12, 14}.
• target (np.ndarray): Single-channel continuous regression target.

---

Usage

from satvision_toa.data_utils.utils_ocean_color import get_dataloaders
from satvision_toa.models.mim import build_mim_model
from satvision_toa.models.decoders.ocean_color_decoder import OceanColorFCNV2point5
from satvision_toa.data_utils.utils_ocean_color import load_config

# Load data
train_dir = "/path/to/chips/ft/chips_ft"
val_dir = "/path/to/chips/ft/val_chips_ft"
train_loader, val_loader, test_loader = get_dataloaders(train_dir, val_dir, num_inputs=14, batch_size=64)

# Build model
config = load_config()
encoder = build_mim_model(config).encoder
model = OceanColorFCNV2point5(swin_encoder=encoder, freeze_encoder=True)

---

Training

Default hyperparameters:

• Epochs: 100
• Batch size: 64
• Optimizer: Adam
• Initial LR: 1e-4 (with decay schedule)

---

Citation

If you use this dataset, please cite the SatVision project:

@software{satvision2024,
  title={SatVision: Foundation Models for Satellite Imagery},
  author={NASA NCCS Data Science Group},
  url={https://github.com/nasa-nccs-hpda/satvision-toa},
  year={2024}
}