How to Use CatBoost to Predict Kharif Crop Yields in Chhattisgarh

CatBoost can be a strong baseline for predicting kharif yields in Chhattisgarh because agricultural datasets combine numerical measurements—rainfall, temperature, soil nutrients and vegetation indices—with categorical fields such as district, crop, soil type and irrigation status. But a useful model depends less on choosing an advanced algorithm than on defining the prediction date, preventing data leakage and validating it against future seasons.

This guide presents a practical workflow for rice, maize, soybean, cotton and other kharif crops. It is designed for analysts, agritech teams, researchers and government programmes working with farm, block or district-level data as of 2026.

Define the prediction problem first

Choose the target, unit of prediction and forecast deadline before collecting features. Possible targets include:

Yield in tonnes per hectare for a specific crop and season.
Production in tonnes, if cultivated area is also modelled reliably.
Yield anomaly compared with a district’s historical average.

A farm-level model needs plot identifiers, sowing dates and management records. A district-level model may use crop area, weather grids, soil data and remote sensing. Do not mix these units without a clear aggregation method: a model trained on district averages cannot be presented as a precise farm-level forecast.

Also define when predictions will be generated. A pre-sowing forecast can use historical climate and soil information; a mid-season forecast can include rainfall accumulated after sowing and satellite indicators; a pre-harvest forecast can use a wider feature set. Every feature must have been available on that forecast date.

For a broader remote-sensing workflow, see satellite-based yield prediction for insurance providers in India. For farm monitoring inputs, automated crop health monitoring systems in India provides useful context.

Assemble Chhattisgarh-specific data

Build one row per prediction unit, crop and season. A useful schema might include district, block, crop, season_year, sown_area, yield, sowing_date and harvest_date.

Potential data sources and feature groups include:

Historical outcomes: crop-cutting experiments, official agricultural statistics, procurement records and carefully quality-checked farm surveys.
Weather: daily or dekadal rainfall, maximum and minimum temperature, humidity, dry-spell length, extreme-rain events and rainfall deviation from the local normal.
Soil: pH, organic carbon, available nitrogen, phosphorus, potassium, texture and drainage.
Management: seed variety, sowing window, fertiliser application, irrigation, crop intensity, pest incidence and tillage practice.
Remote sensing: NDVI, EVI, land-surface temperature, water indices and cloud-free seasonal composites.
Location: district, block, elevation, agro-climatic zone and distance to irrigation or markets.

Chhattisgarh’s rainfall and cropping conditions vary across districts, so statewide averages can hide important local patterns. Retain district and block identifiers, but check whether they represent genuine agronomic differences or simply memorise the target. Keep a data dictionary recording units, collection dates, spatial resolution and missing-value rules.

Engineer features without leaking future information

Aggregate weather relative to sowing or a fixed calendar window. Examples include rainfall during establishment, cumulative rainfall during the vegetative phase, number of dry days, heat-stress days and rainfall concentration. For satellite data, calculate trends and phase-specific summaries rather than using a single maximum value that may be affected by clouds or an unusually noisy observation.

Useful derived features include:

Rainfall deviation from a district’s historical baseline.
Rolling seven-, 14- and 30-day rainfall totals.
Growing degree days, where appropriate for the crop.
NDVI change between early and peak growth stages.
Irrigated-area share and crop-specific sown-area change.
Interactions between soil properties, rainfall and crop type.

Do not use final harvested area, post-harvest production, revised yield estimates or satellite observations captured after the forecast date. Impute only with information available in the training period. If a variable is unavailable for a new season, excluding it is better than silently substituting a future-derived value.

Train CatBoost for regression

Install the core packages:

pip install catboost pandas scikit-learn joblib

A minimal training structure is:

import pandas as pd
from catboost import CatBoostRegressor
from sklearn.metrics import mean_absolute_error, mean_squared_error

features = [
    "district", "block", "crop", "season_year",
    "rainfall_30d", "rainfall_deviation", "dry_days",
    "ndvi_peak", "soil_ph", "soil_n", "irrigated_share"
]
cat_features = ["district", "block", "crop"]

train = pd.read_csv("kharif_yield_train.csv")
valid = pd.read_csv("kharif_yield_valid.csv")

model = CatBoostRegressor(
    loss_function="RMSE",
    eval_metric="MAE",
    iterations=2000,
    learning_rate=0.04,
    depth=7,
    l2_leaf_reg=8,
    random_seed=42,
    od_type="Iter",
    od_wait=100,
    verbose=200
)

model.fit(
    train[features], train["yield_t_ha"],
    cat_features=cat_features,
    eval_set=(valid[features], valid["yield_t_ha"]),
    use_best_model=True
)

pred = model.predict(valid[features])
mae = mean_absolute_error(valid["yield_t_ha"], pred)
rmse = mean_squared_error(valid["yield_t_ha"], pred) ** 0.5
print({"MAE": mae, "RMSE": rmse})

CatBoost can process categorical columns directly, but convert missing categorical values to a consistent string such as unknown. Ensure numeric columns are genuinely numeric and that target values use consistent units. A log-transformed target may help when yields have a long right tail, but report results back in tonnes per hectare.

Validate by season, not just random rows

A random 80:20 split is often misleading. Rows from the same district and season can be highly similar, allowing the model to learn conditions that would not be known in a future forecast. Use a time-based design instead:

Train on earlier seasons and validate on the next season.
Hold out the most recent season as a final test set.
If farm-level records are used, group by farm or block where repeated observations could leak information.
Compare performance for rice, maize and other crops separately.

Report MAE, RMSE, mean bias and percentage errors carefully. Include a simple baseline, such as the district’s historical mean yield. A model that does not beat this baseline consistently may not justify operational complexity. Also report uncertainty: prediction intervals, ensemble spread or error bands by crop and district are more useful to decision-makers than a single point estimate.

Explain errors and assess fairness

Use CatBoost feature importance or SHAP values to understand whether rainfall, vegetation, soil or location is driving predictions. Explanation is not proof of causality. Investigate suspicious patterns—for example, a model relying almost entirely on district identity may be memorising historical reporting differences.

Create an error dashboard showing actual versus predicted yield by district, crop, season, irrigation status and farm size. Check whether performance deteriorates for rainfed farms, tribal regions, smallholders or locations with sparse observations. Flag low-confidence predictions rather than presenting them as precise advice.

For production teams, implementing scalable ML pipelines for predictive analytics offers a useful framework for versioning data, monitoring features and scheduling retraining.

Deploy a useful forecast service

Save the model and its feature contract:

model.save_model("kharif_catboost_yield.cbm")

A FastAPI service, scheduled batch job or dashboard can return the predicted yield, forecast date, input completeness and uncertainty range. Validate incoming data types, reject impossible values and log the model version with every prediction. Store the exact weather and satellite snapshot used so results remain auditable.

Retrain after each season only after quality checks. Monitor feature drift, missingness, changes in crop varieties and shifts in reporting practices. If remote-sensing availability changes because of cloud cover or a new processing pipeline, performance can fall even when the farm system itself has not changed.

The forecast should support—not replace—field validation and local agronomy. Pair model outputs with advisories, insurance workflows, procurement planning or resource allocation, and communicate limitations plainly. Teams looking to connect prediction with agronomic action can also review how to improve crop yield with AI in India.

Practical checklist

Before using a Chhattisgarh kharif yield model operationally, confirm that you have:

A clearly defined crop, geography, target and forecast date.
Historical data with documented units and collection methods.
Leakage-safe temporal validation and a naive baseline.
Crop- and district-level error analysis.
Missing-data, drift and retraining procedures.
Explanations and uncertainty estimates for users.
A deployment log linking every forecast to its data and model version.

CatBoost is a capable, practical starting point, not a guarantee of accuracy. Better labels, consistent seasonal data and honest validation will usually produce more value than adding arbitrary depth or iterations to the model.

Apply for AI Grants India