AgMultida Water Stress Proxy Dataset (data/processed_real)
Pipeline validation and first-pass training for rice water-stress detection and water-saving irrigation recommendation.
- Canonical training root:
data/processed_real/ - Final manifest:
data/processed_real/sample_manifest.csv - Historical expansion is expected to change sample counts over time.
- Labels: continuous proxy stress label
[0, 1] - Primary training metrics should be continuous (
loss,mae,rmse); binary metrics are diagnostics only.
- Sentinel-2 RGB+NIR image patch:
[4, 224, 224] - Weather-derived sensor sequence:
[48, 8] - Weather context:
[6] - Modality mask:
[3] - Label:
[1]
Labels are proxy labels based on NDVI anomaly, ET deficit, rain relief, and heat penalty. They are not field-measured stress labels and should not be presented as agronomist-verified ground truth.
sensor_seq is weather-derived proxy data, not telemetry from physical field devices. It is appropriate for Phase 1 pipeline validation but not for claims about field-sensor performance.
Zones are synthetic validation polygons, not audited farm boundaries.
Safe claim: multimodal proxy-label regression pipeline for water-stress scoring and irrigation recommendation.
Unsafe claims:
- field-ready agronomic accuracy
- rice blast detection
- brown planthopper detection
- nitrogen deficiency detection
- physical sensor telemetry validation
The manifests and .npy tensors must pass integrity checks before full training. Manifest loader now expects source_status == PASS and can optionally validate checksums. Run scripts/check_data_integrity.py --data-dir data/processed_real before VPS launch.