krishikarm 0.1.0

Crop distress prediction from satellite and weather data using a cross-attention transformer (KisanNet v3).

Krishikarm

Crop distress prediction from satellite and weather data.

Krishikarm uses KisanNet v3, a cross-attention transformer trained on 82,000+ real satellite observations across 127 Indian districts, to predict crop health risk in real time.

Installation

pip install krishikarm

Quick Start

from krishikarm import Predictor

predictor = Predictor()

result = predictor.predict({
    "lat": 12.97,
    "lon": 77.59,
    "T2M": 32,
    "RH2M": 65,
    "PREC": 2.5,
    "SOLAR": 18,
    "WIND": 3,
    "state": "KA",
    "irrig": "borewell",
    "crops": ["rice"],
})

print(result)
# {
#     "distress_score": 0.12,
#     "intervention_days": 26.4,
#     "risk_class": 0,
#     "risk_label": "Healthy"
# }

Model Architecture

KisanNet v3 is a cross-attention transformer with:

• 42 input features derived from weather, soil, atmosphere, and location data
• State and irrigation embeddings for regional context
• 4-layer, 8-head cross-attention fusion between satellite and context encoders
• Three prediction heads: distress score (0-1), intervention window (days), risk class (5 levels)
• 96.8% accuracy on held-out test data

Input Features

Source	Features
Weather	Temperature, humidity, precipitation, wind, cloud cover, dew point
Solar	Shortwave radiation, UV index, longwave radiation
Soil	Moisture (3 depths), temperature (2 depths), clay, sand, SOC, pH
Atmosphere	PM2.5, PM10, dust, aerosol optical depth
Derived	NDVI proxy, VPD, GDD, thermal range
Location	Latitude, longitude, elevation, day-of-year encoding

Risk Classes

Class	Label	Distress Range
0	Healthy	0.00 - 0.15
1	Watch	0.15 - 0.30
2	Alert	0.30 - 0.50
3	Critical	0.50 - 0.75
4	Emergency	0.75 - 1.00

Data Sources

All training data was collected from free, open APIs:

• NASA POWER -- Solar radiation, temperature, humidity
• Open-Meteo -- Weather forecasts, soil moisture, UV, air quality
• SoilGrids -- Soil composition (clay, sand, SOC, pH)
• Open-Elevation -- Terrain altitude

No proprietary data or paid APIs were used.

API Reference

Predictor

Predictor(weights_path=None, device=None)

• weights_path -- Path to a .pth checkpoint. Defaults to the bundled v3 weights.
• device -- "cpu" or "cuda". Defaults to CUDA when available.

Methods

• predict(observation: dict) -> dict -- Run prediction on a single observation.
• predict_batch(observations: list[dict]) -> list[dict] -- Batch prediction.
• info() -> dict -- Return model metadata.

KisanNetV3

The raw PyTorch model class, for advanced users who want to integrate the architecture into their own training pipeline.

Requirements

• Python >= 3.9
• PyTorch >= 2.0
• NumPy >= 1.24

License

MIT