Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
1.26k
4.5k
End of preview. Expand in Data Studio

AgriSense Fusion: High Barind Tract Earth Observation & Agroclimatology Dataset

NASA Space Apps Challenge 2026 — Challenge: Field Shift: Adapting Farms with NASA Data
Primary Study Site: Godagari Upazila, Rajshahi (AEZ-26 High Barind Tract · 24.4690°N, 88.3300°E)
Source Code Repository: ASHIK311/agrisense-fusion


📦 Dataset Contents (~2.80 GB)

This repository hosts the large multi-temporal geospatial GeoTIFF rasters, trained Random Forest binary classifiers, and multi-decade NASA POWER agroclimatology tables that accompany the lightweight codebase on GitHub.

1. satellite_gis_barind/ (Sentinel-2 L2A & Spectral Indices, ~2.78 GB)

  • data/raw/: 12 multi-year Sentinel-2 Level-2A acquisition scenes (2022-10-21 to 2025-04-05), bands B02, B03, B04, B08, B11, B12.
  • data/processed/: 12 clipped and cloud-masked composite rasters (*_godagari_preprocessed.tif).
  • data/labels/: ESA WorldCover baseline masks (esa_worldcover_godagari.tif, esa_worldcover_godagari_polygon_utm.tif).
  • indices/ & exports/geotiff/: Precomputed spectral indices (NDVI, NDWI McFeeters, NDMI Gao, EVI) across all 12 dates.
  • maps/classification/: Land Use / Land Cover (LULC) classified rasters, prediction confidence maps, and spatial holdout error maps.
  • models/: Trained scikit-learn / Random Forest .pkl model artifacts (RandomForest_21Feat_Deep and baseline checkpoints).
  • reports/: High-resolution spatial figures and the compiled 34.8 MB PDF scientific report (final_member1_report.pdf).

2. agronomy_national_climate/ (~15 MB)

  • nasa_power_monthly_all_upazilas_1981_2024.csv: Long-term monthly NASA POWER agroclimatological time-series across Bangladesh upazilas (1981–2024).

📊 Verified Performance & Honest Scientific Limits

  • Model: RandomForest_21Feat_Deep (21 temporal/spectral features, 4 broad LULC classes: Cropland, Tree Cover, Shrub/Grass/Bare, Water/Built).
  • Block C Spatial Holdout ($N = 6,000$ synthetic holdout samples):
    • Overall Accuracy: 85.77% (5,146 / 6,000)
    • Balanced Accuracy: 84.02%
    • Macro-F1: 0.8257 | Weighted-F1: 0.8589
    • Pure Interior Centroid Accuracy: 90.64% (3,639 / 4,015)
    • Mixed Boundary Accuracy: 75.92% (1,507 / 1,985)
  • Spatial 5-Fold CV ($N = 20,000$ across Blocks A & B): 84.63% ± 3.81% (0.8342 ± 0.0194 Macro-F1).
  • Important Scientific Disclosures:
    • Block C is a southern latitudinal development holdout partition without an inter-block spatial buffer zone—not an untouched external field survey.
    • Fine-grained crop species classification (e.g., Boro rice vs. Wheat vs. Mustard at parcel scale) remains explicitly gated/blocked (HTTP 422) pending field-surveyed DAE cadastral crop polygons.

⬇️ Quick Download into GitHub Repo

If you cloned ASHIK311/agrisense-fusion and want to pull these heavy rasters locally:

pip install huggingface_hub
python scripts/download_dataset.py

Or directly via the Hugging Face CLI:

hf download ashik1236/nasatest --repo-type dataset --local-dir ./research/member_01_satellite_gis/hf_data

📜 Attribution & License

  • Earth Observation Data: Copernicus Sentinel-2 L2A (ESA), ESA WorldCover 10m, NASA POWER (LaRC).
  • Code & Dataset Packaging: Team AgriSense Fusion (Dhaka, Bangladesh) — Released under the MIT License.
Downloads last month
239