Apply for AI Grants India

Sikkim’s large-cardamom growers work with steep terrain, fragmented farms, and highly variable rainfall. A useful forecast is therefore not just a prediction of tomorrow’s temperature. It should help answer practical questions: Will the coming week be too wet for field work? Is humidity likely to remain high enough for fungal disease? Should irrigation, shade management, drainage, or harvesting plans change?

Prophet can provide a strong baseline for these questions when it is trained on reliable, local time-series data. It is not a replacement for an IMD warning, agronomist, or field observation, but it can turn historical observations into a repeatable planning tool.

What Prophet can and cannot predict

Prophet is an open-source time-series forecasting library originally developed at Facebook and now maintained as an independent open-source project. It models:

• Long-term trend in a measured variable
• Repeating seasonality, such as monsoon and winter patterns
• Known events or interventions through additional regressors or holiday-style features
• Uncertainty around each forecast

For Sikkim cardamom, build separate models for variables such as:

• Daily or weekly rainfall
• Minimum and maximum temperature
• Relative humidity
• Soil moisture, if sensors are available
• Leaf-wetness duration or disease observations

Prophet forecasts the variable supplied in the y column. It does not automatically forecast cardamom yield from weather. To estimate yield, you need a second model or a carefully designed feature-based workflow that combines forecast weather with crop, soil, and management data.

Choose data that represents the farm

The largest source of error will usually be data mismatch, not Python syntax. Weather in Sikkim changes sharply with elevation, slope, aspect, and valley exposure. A station several kilometres away may not represent a cardamom plot.

Start with the following data sources:

• IMD station or gridded rainfall and temperature data
• Automatic weather stations operated by agricultural institutions or local projects
• Farm-level sensors for humidity, temperature, soil moisture, and rainfall
• Crop records covering planting, flowering, disease, harvest, and yield
• Elevation, slope, aspect, shade-tree cover, and plot location

Keep the source, unit, timestamp, elevation, and missing-data treatment for every variable. Record whether rainfall is measured from midnight to midnight, in a rolling 24-hour window, or using another convention. Consistent definitions matter when comparing forecasts with field outcomes.

For geospatial and sensor-heavy projects, maintain a clean data pipeline before adding complex AI. Teams building broader agricultural systems may also benefit from practices described in how to build computer vision models on GitHub, especially when combining weather data with disease or canopy images.

Prepare the Prophet dataset

Prophet expects a dataframe with two required columns:

• ds: a date or timestamp
• y: the numeric value being forecast

A minimal daily rainfall dataset might look like this:

import pandas as pd
from prophet import Prophet

weather = pd.read_csv("sikkim_weather.csv")
weather["ds"] = pd.to_datetime(weather["date"])
weather["y"] = pd.to_numeric(weather["rainfall_mm"], errors="coerce")
weather = weather[["ds", "y"]].sort_values("ds")
weather = weather.drop_duplicates("ds")

Before fitting the model:

• Convert all measurements to consistent units.
• Identify impossible values, such as negative rainfall or humidity above 100%.
• Distinguish a true zero from a missing observation.
• Aggregate hourly readings to daily or weekly values when appropriate.
• Keep a record of imputation instead of silently filling gaps.

For rainfall, a weekly model may be more stable than a daily model because daily precipitation is intermittent and often contains many zero values. Temperature and humidity can generally be modelled daily if the station record is sufficiently complete.

Train a location-aware baseline

Install the current package with pip install prophet. The modern import is from prophet import Prophet; older examples using fbprophet are outdated.

from prophet import Prophet

model = Prophet(
    yearly_seasonality=True,
    weekly_seasonality=False,
    daily_seasonality=False,
    seasonality_mode="multiplicative",
    interval_width=0.90
)

model.fit(weather)
future = model.make_future_dataframe(periods=30, freq="D")
forecast = model.predict(future)

result = forecast[["ds", "yhat", "yhat_lower", "yhat_upper"]]
print(result.tail(30))

Do not assume that every default is suitable for Sikkim. Compare additive and multiplicative seasonality, and test whether weekly seasonality has any agricultural meaning. Add custom seasonal patterns only when the data supports them:

model.add_seasonality(
    name="monsoon",
    period=365.25,
    fourier_order=8
)

Avoid duplicating yearly seasonality without testing it. A model that fits historical monsoon peaks perfectly may still fail during an unusual rainfall year.

Add cardamom-relevant predictors carefully

Prophet can include external regressors, but those regressors must also be available for the future period. Suitable candidates may include:

• Elevation or plot-specific fixed characteristics
• Forecast temperature when predicting humidity or disease risk
• Soil moisture readings
• A monsoon-season indicator
• Irrigation or shade-management events

For example:

train = weather.merge(sensor_data, on="ds", how="left")
model = Prophet(yearly_seasonality=True)
model.add_regressor("soil_moisture")
model.fit(train[["ds", "y", "soil_moisture"]])

You must provide soil_moisture in the future dataframe too. If future values are unknown, use a separate forecast, a scenario, or omit the regressor. Otherwise, the model will not produce a defensible forecast.

For crop-yield prediction, create features such as cumulative rainfall over the previous 7, 14, and 30 days; consecutive wet days; temperature range; and humidity persistence. Then compare Prophet with tree-based models or regression. The objective should be a farm decision, not simply a visually attractive chart.

Validate with walk-forward testing

Random train-test splits leak future information into the past and are unsuitable for time series. Use rolling, walk-forward validation:

1. Train on the earliest period.
2. Forecast the next 7, 14, or 30 days.
3. Compare the forecast with observations.
4. Expand the training window and repeat.

Report MAE and RMSE for temperature, while rainfall may need MAE, bias, wet-day accuracy, and threshold metrics such as the ability to identify rainfall above 25 mm. For farmer-facing alerts, calibration matters: a 90% prediction interval should contain the actual value approximately 90% of the time.

Compare Prophet against simple baselines, including yesterday’s value, seasonal averages, and a moving average. If Prophet does not outperform these baselines consistently, do not deploy it as the primary decision tool.

Turn forecasts into farm actions

A forecast becomes useful when linked to clear thresholds and local advice. Examples include:

• Delay spraying when heavy rain is likely soon after application.
• Inspect drainage after several consecutive wet days.
• Increase disease scouting when humidity and leaf wetness remain elevated.
• Plan harvesting and transport around predicted dry windows.
• Use uncertainty bands to trigger conservative action when the forecast is ambiguous.

Present forecasts in Nepali or another locally appropriate language, with a simple explanation of confidence and the last data-update time. A mobile dashboard should show the plot location, forecast horizon, observed values, and alert thresholds—not only a model score. If the system later includes leaf or disease images, evaluate its visual pipeline separately from the weather model; model deployment considerations are covered in AI model optimisation for mobile devices.

Key limitations and safeguards

Prophet is not a numerical weather prediction system. It learns patterns from historical observations and may miss sudden cloudbursts, landslides, long dry spells, or climate-regime shifts. Sikkim’s microclimates make station selection especially important.

Use IMD advisories and local agricultural expertise for safety-critical decisions. Monitor data drift, retrain after adding a reliable season of observations, and keep an audit trail of model versions. Protect farmer and location data, particularly when combining plot coordinates with production records.

A practical implementation checklist

• Define the decision the forecast must support.
• Select the nearest representative station or install plot sensors.
• Build separate datasets for each target variable.
• Clean timestamps, units, missing values, and outliers.
• Establish seasonal and naive baselines.
• Train Prophet with conservative seasonality settings.
• Validate using walk-forward backtesting.
• Convert forecast thresholds into agronomic actions.
• Show uncertainty and data freshness to users.
• Review results with Sikkim growers and agricultural specialists.

Prophet is most valuable here as a transparent, maintainable baseline. With location-aware data, honest validation, and clear farm thresholds, it can support better cardamom planning without overstating what a statistical forecast can know.