Datasets:

SanuthK
/

AgroGate_SAR2NDVI

AgroGate_SAR2NDVI provides pixel-aligned pairs of C-band SAR image chips (VV, VH) from Sentinel-1 and corresponding NDVI targets computed from Sentinel-2. Each pair is cropped around paddy areas in Sri Lanka and saved as small GeoTIFF tiles, enabling supervised learning to predict optical vegetation indices from radar, study SAR–optical relationships, and build cloud-robust vegetation monitoring models.

Geography: Sri Lanka (paddy regions)
Years available: 2023, 2022, 2021 (more years coming)
Inputs (X): Sentinel-1 GRD, bands VV, VH
Targets (Y): NDVI derived from Sentinel-2 (B8, B4)
Tile size: ~128×128 px (10 m)
Format: GeoTIFF (2-band SAR, 1-band NDVI)
Pairing: Filenames match across SAR/ and NDVI/ (e.g., 2023_yala_0.tif)

Motivation

Learn radar→optical mappings to predict NDVI from SAR when optical is cloud-obscured.
Support crop monitoring and digital twin applications for Sri Lankan paddy.
Provide an easy, year-wise downloadable dataset for research and ML benchmarks.

Data Sources & Processing

Sentinel-1 GRD (IW): filtered by ROI and date (±3 days around Sentinel-2 date), bands VV, VH, clipped to tiles.
Sentinel-2 SR (Harmonized): low-cloud scenes, NDVI = (B8 − B4) / (B8 + B4), clipped to tiles.
Sampling: points taken from paddy masks; each point yields multiple date pairs within the paddy season.
Seasons: Maha (Oct–Mar), Yala (Apr–Sep).
Pairing rule: each NDVI date uses the nearest available Sentinel-1 acquisition within ±3 days.

Contains modified Copernicus Sentinel data.

Repository Structure

AgroGate_SAR2NDVI/
├── 2023/
│   ├── SAR/
│   │   ├── 2023_yala_0.tif        # 2 bands: [VV, VH]
│   │   ├── 2023_yala_1.tif
│   │   └── ...
│   ├── NDVI/
│   │   ├── 2023_yala_0.tif        # 1 band: NDVI
│   │   ├── 2023_yala_1.tif
│   │   └── ...
│   └── Plots/                     # optional quick-look PNGs
└── (future years)/                # 2022, 2024, ...

Each pair is identified by a shared stem (e.g., 2023_yala_0) present in both SAR/ and NDVI/.

File Format & Bands

SAR tiles (`SAR/*.tif`)

Type: GeoTIFF
Bands:
1. VV (float32)
2. VH (float32)

NDVI tiles (`NDVI/*.tif`)

Type: GeoTIFF
Bands:
1. NDVI (float32, range roughly −1…1)

All tiles are projected and cropped to identical extents per pair.

How to Use

Load year-specific data with `datasets`

from datasets import load_dataset

# Download only the 2023 subset
ds = load_dataset("SanuthK/AgroGate_SAR2NDVI", data_dir="2023")

If you kept the raw folder structure, datasets will primarily fetch the files; pairing logic is typically handled in your ML dataloader (see below).

Simple PyTorch dataset for paired loading

import os, rasterio, torch
from torch.utils.data import Dataset

class SarNdviDataset(Dataset):
    def __init__(self, sar_dir, ndvi_dir):
        self.sar_dir = sar_dir
        self.ndvi_dir = ndvi_dir
        self.ids = sorted([f[:-4] for f in os.listdir(sar_dir) if f.endswith(".tif")
                           and os.path.exists(os.path.join(ndvi_dir, f))])

    def __len__(self): return len(self.ids)

    def __getitem__(self, i):
        pid = self.ids[i]
        with rasterio.open(os.path.join(self.sar_dir,  f"{pid}.tif")) as src:
            sar = src.read().astype("float32")      # (2, H, W)
        with rasterio.open(os.path.join(self.ndvi_dir, f"{pid}.tif")) as src:
            ndvi = src.read(1).astype("float32")    # (H, W)
        return torch.from_numpy(sar), torch.from_numpy(ndvi).unsqueeze(0)

Usage:

ds2023 = SarNdviDataset(
    sar_dir="/path/to/AgroGate_SAR2NDVI/2023/SAR",
    ndvi_dir="/path/to/AgroGate_SAR2NDVI/2023/NDVI"
)
x, y = ds2023[0]  # x: (2,H,W), y: (1,H,W)

Splits

No predefined splits are provided. Recommended strategies:

Spatial split: divide by geographic tiles/points to avoid leakage.
Temporal split: split by dates (e.g., early vs. late season).
Seasonal split: train on Yala, validate on Maha, or vice versa.

Intended Uses

Train models to predict NDVI from SAR (cloud-robust vegetation monitoring).
Study relationships between SAR backscatter (VV/VH) and vegetation vigor.
Downstream tasks: yield proxies, crop phenology tracking, missing-data imputation.

Limitations

Cloud dependency for NDVI: tiles exist only where low-cloud S2 scenes were available.
Temporal gap: SAR is within ±3 days of the NDVI date; not strictly same-day.
Region-specific: focused on Sri Lankan paddy systems—generalization to other regions/crops may require adaptation.

Ethical Considerations

Remote sensing over agricultural areas; no PII.
Use responsibly and avoid claims beyond what the data supports (e.g., exact yield without calibration).

Licensing

Dataset license: CC-BY-4.0
Upstream data: Contains modified Copernicus Sentinel data (2015–present) under the Copernicus free and open data policy.
Please retain attribution in any derivative works.

Citation

If you use AgroGate_SAR2NDVI in your research, please cite:

@dataset{AgroGate_SAR2NDVI_2025,
  title   = {AgroGate_SAR2NDVI: Paired Sentinel-1 SAR and Sentinel-2 NDVI Tiles for Sri Lankan Paddy},
  author  = {Wanniarachchi S.K. and Weerakoon D.S.K},
  year    = {2025},
  publisher = {Hugging Face Datasets},
  note    = {Contains modified Copernicus Sentinel data},
  howpublished = {\url{https://huggingface.co/datasets/SanuthK/AgroGate_SAR2NDVI}}
}

Changelog

2025-08-27: Initial public release with 2023 season tiles (Yala/Maha).
Future: add 2022, 2024+, richer metadata (dates, ROI coords), and predefined splits.