Apply for AI Grants India

Why chickpea yield prediction matters in Madhya Pradesh

Madhya Pradesh is one of India’s most important chickpea-producing states, but yields vary sharply across districts and seasons. Rainfall distribution, terminal heat, soil moisture, sowing dates, irrigation access, pest pressure, and input decisions all influence the final harvest. A useful model should therefore do more than produce a number: it should provide an auditable estimate early enough to support procurement, crop insurance, advisories, and farm planning.

A feedforward neural network (FNN) is a sensible starting point when the dataset contains structured observations—such as weather summaries, soil measurements, crop-management variables, and historical yield—and the target is a continuous value such as tonnes per hectare. It is not automatically superior to linear regression, random forests, or gradient boosting. The right choice depends on data volume, feature quality, validation design, and how much interpretability stakeholders require.

Define the prediction task before collecting data

Specify four things upfront:

• Target: yield in kg/ha or tonnes/ha, recorded consistently across years and locations.
• Prediction horizon: pre-sowing, mid-season, or near-harvest. The available features must exist by the forecast date.
• Spatial unit: district, block, village, field, or grid cell. Avoid mixing these units without a clear aggregation method.
• Decision use: crop insurance, procurement, extension advice, seed planning, or research.

For example, a district-season model might predict yield 30 days before harvest using rainfall accumulated after sowing, maximum temperature, vegetation indices, soil properties, and reported management practices. A field-level model needs much finer labels and is harder to validate. If satellite data is available, compare the FNN against approaches described in satellite-based yield prediction for insurance providers.

Build a Madhya Pradesh-focused dataset

Useful inputs include:

• Historical yield: official district or block statistics, digitised crop-cutting experiments, and carefully verified farm records.
• Weather: daily rainfall, minimum and maximum temperature, humidity, solar radiation, and dry-spell indicators. Convert daily observations into agronomically meaningful windows rather than feeding every raw value blindly.
• Soil: pH, organic carbon, available nitrogen, phosphorus, potassium, texture, and drainage. Keep sampling depth and laboratory methods consistent.
• Crop management: sowing date, variety, seed rate, fertiliser application, irrigation events, weed control, and preceding crop.
• Remote sensing: NDVI or other vegetation measures, surface temperature, and moisture proxies, aligned to the crop calendar.
• Geography: district, elevation, soil zone, and irrigation classification. Encode categorical variables carefully; do not assign arbitrary numeric meanings to districts.

Record data provenance, units, collection dates, and missingness. In India, administrative boundaries and reporting practices can change, so maintain a versioned geographic crosswalk. A spreadsheet that lacks source and timestamp fields will become a liability during validation.

Prevent leakage and prepare the features

Agricultural datasets often contain leakage: information that would not have been known at the stated prediction date. A final harvest statistic, late-season rainfall, or a post-harvest survey field must not enter an early-season model. Write a feature-availability rule for every column.

Use this preparation sequence:

1. Remove duplicate district-season records and reconcile conflicting yield values.
2. Standardise units, dates, district names, and missing-value codes.
3. Impute missing values using training data only. Median imputation is a defensible baseline; weather gaps may require station interpolation.
4. Create crop-stage features such as rainfall during sowing, vegetative, flowering, and pod-filling windows.
5. Add biologically plausible interactions, such as heat during flowering or rainfall following a dry spell.
6. Scale continuous variables using a scaler fitted only on the training partition.
7. Keep a separate data dictionary and preprocessing pipeline so the same transformations are used in production.

For repeatable deployment, structure preprocessing and training as an implementing scalable ML pipeline for predictive analytics, even if the first prototype runs on a laptop.

Choose a validation strategy that reflects the real forecast

A random 80/20 split can make results look stronger than they will be in practice because neighbouring districts and adjacent years share weather patterns. Prefer validation that mirrors deployment:

• Time-based split: train on earlier seasons and test on later seasons.
• Leave-one-district-out: test geographic transfer to an unseen district.
• Group-based folds: keep all records from the same district-season in one fold.
• Rolling-origin evaluation: repeatedly train on the past and predict the next season.

Fit the scaler, imputers, feature selectors, and model inside each training fold. Report performance by district, season, irrigation category, and yield range—not just one overall score. Use MAE for an understandable average error, RMSE to penalise large misses, and R² as a supplementary measure. Also compare against a historical mean, linear regression, random forest, and gradient boosting. A neural network should earn its place through out-of-sample performance and operational value.

Build a compact feedforward neural network

For a tabular dataset, begin with a small architecture: one or two hidden layers, ReLU activations, and a single linear output for yield. Larger networks can memorise a modest agricultural dataset rather than learn general patterns. Practical controls include early stopping, dropout, L2 regularisation, and a learning-rate schedule. Explore architectures systematically rather than selecting the best result from one lucky split; the principles behind customizable neural network architectures for beginners are useful here.

import pandas as pd
import tensorflow as tf
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import mean_absolute_error, mean_squared_error

# One row per district-season; target is yield_t_per_ha
df = pd.read_csv("chickpea_yield_mp.csv")
features = ["rain_sowing", "rain_flowering", "tmax_flowering",
            "soil_ph", "soil_n", "ndvi_pod_filling"]
X = df[features]
y = df["yield_t_per_ha"]

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)

model = tf.keras.Sequential([
    tf.keras.layers.Input(shape=(X_train.shape[1],)),
    tf.keras.layers.Dense(64, activation="relu"),
    tf.keras.layers.Dropout(0.15),
    tf.keras.layers.Dense(32, activation="relu"),
    tf.keras.layers.Dense(1)
])
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3),
              loss="mse", metrics=[tf.keras.metrics.MeanAbsoluteError()])

stop = tf.keras.callbacks.EarlyStopping(
    monitor="val_loss", patience=20, restore_best_weights=True
)
model.fit(X_train, y_train, validation_split=0.2,
          epochs=300, batch_size=16, callbacks=[stop], verbose=0)

pred = model.predict(X_test, verbose=0).ravel()
print("MAE:", mean_absolute_error(y_test, pred))
print("RMSE:", mean_squared_error(y_test, pred) ** 0.5)

For a production study, replace the demonstration split with grouped or time-based cross-validation. Save the trained scaler and model together, and reject inputs outside the training range unless the system explicitly handles extrapolation.

Interpret, stress-test, and communicate predictions

Farmers and public agencies need to know why a forecast changed. Use permutation importance, partial-dependence analysis, or SHAP with appropriate care. Check whether the model behaves sensibly when rainfall, heat, or soil fertility changes. Test missing weather stations, delayed satellite observations, and district names not present during training.

A prediction should include an uncertainty estimate, not only a point value. Ensembles across several trained networks, quantile models, or conformal prediction can provide an interval. Communicate uncertainty in operational terms—for example, an expected yield range and the conditions that could move it outside that range. Avoid claiming field-level precision when labels exist only at district level.

Deployment checklist for 2026 projects

Before using the model for decisions:

• Establish a baseline and document whether the FNN improves it.
• Validate on a genuinely later season or unseen geography.
• Audit errors by crop variety, irrigation status, soil zone, and district.
• Monitor drift in weather distributions, sowing dates, satellite coverage, and reporting quality.
• Store model versions, feature snapshots, predictions, and corrections.
• Obtain consent and protect farmer-level records where applicable.
• Give agronomists a review path; a forecast should support, not replace, local expertise.

The most valuable system is often a modest model connected to reliable data, clear uncertainty, and timely advisories. FNNs can capture non-linear relationships in chickpea production, but disciplined data design and validation determine whether those patterns are useful in Madhya Pradesh.