Madhya Pradesh is one of India’s important garlic-producing states, with output shaped by district, season, irrigation, soil, weather, seed choice, disease pressure, and market-linked farm decisions. A useful forecast is not simply a high-accuracy number: it should help farmers, aggregators, insurers, processors, and government teams plan procurement, storage, inputs, and risk mitigation.
This guide explains how to use radial basis function networks to predict garlic production in Madhya Pradesh. The focus is a reproducible workflow that works with limited agricultural datasets and can be improved as more farm- and satellite-level observations become available.
What an RBF network does
A radial basis function network (RBFN) is a feed-forward neural network suited to regression and function approximation. It learns how combinations of inputs—such as rainfall, temperature, cultivated area, soil properties, and irrigation access—relate to a continuous target such as yield or total production.
An RBFN usually has three parts:
- Input layer: Receives engineered features for a district, crop season, or farm.
- Radial basis layer: Measures how close each observation is to learned centres, commonly using Gaussian functions.
- Output layer: Combines those activations to estimate yield or production.
For an observation x and centre c, a Gaussian basis function can be written as:
φ(x) = exp(-||x - c||² / 2σ²)
Here, c is a centre and σ controls the width of the function. The output is a weighted combination of these basis functions. Unlike a simple linear model, an RBFN can represent local, non-linear patterns—for example, the way excess rainfall affects garlic differently from moderate rainfall.
RBFNs are valuable when the dataset is structured, the target relationship is non-linear, and training data is not large enough to justify a very deep model. They should still be compared with strong baselines such as linear regression, random forest, gradient boosting, and a multi-layer perceptron.
Define the prediction target first
Do not combine different forecasting problems under one label. Choose one target and align every input to the forecast date:
- Yield: tonnes per hectare, useful for agronomic and farm-level decisions.
- Production: total tonnes, generally calculated as area multiplied by yield.
- Area: hectares planted, useful for early-season supply forecasts.
- District or state output: aggregated production for procurement and market planning.
For most applications, predict yield first and calculate production separately. This prevents a model from confusing a rise in cultivated area with improved crop performance. If area data is reliable, production can be estimated as:
predicted production = predicted yield × cultivated area
Record the unit, season, district, and forecast horizon for every row. A model trained on district-level annual data should not be presented as a farm-level, week-ahead forecast.
Assemble Madhya Pradesh data
Create a panel dataset with one row per district-season or block-season. Candidate districts may include major garlic-growing areas such as Mandsaur, Neemuch, Ratlam, Ujjain, Indore, and surrounding production belts, but the model should use the actual coverage and labels available rather than assume every district behaves alike.
Useful feature groups include:
- Historical production: yield, sown area, harvested area, and lagged production.
- Weather: minimum and maximum temperature, rainfall totals, rainy-day counts, humidity, heat events, and dry spells.
- Soil: pH, organic carbon, electrical conductivity, texture, nitrogen, phosphorus, potassium, and drainage indicators.
- Farm practices: sowing date, seed variety, seed rate, irrigation type, fertiliser application, pesticide use, and crop rotation.
- Remote sensing: vegetation indices, surface temperature, soil moisture proxies, and phenology metrics from satellite imagery.
- Operational context: irrigation availability, pest alerts, input access, storage capacity, and local market conditions.
Use official agricultural statistics, state and district records, weather stations, validated gridded weather products, soil surveys, farmer surveys, and satellite data. A satellite-based yield prediction workflow for Indian insurers offers a useful reference for combining remote sensing with ground observations.
Prepare the dataset without leakage
Agricultural datasets often contain missing values, inconsistent boundaries, and measurements collected after harvest. These issues can make a model appear accurate while making it unusable in practice.
Follow this sequence:
1. Standardise geography: Map historical district names to stable codes and document boundary changes.
2. Align timing: Include only information that would have been available at the intended prediction date.
3. Handle missingness: Use agronomically sensible imputation, add missing-value indicators where useful, and retain a missing-data log.
4. Remove duplicates: Check repeated survey entries, inconsistent units, and impossible values.
5. Transform skewed variables: Consider log transformation for production, area, or rainfall where distributions are highly uneven.
6. Scale features: Standardise numeric inputs before calculating RBF distances; otherwise large-unit variables dominate similarity.
7. Encode categories: Represent soil class, irrigation type, or variety using one-hot encoding or carefully designed agronomic groupings.
Use a time-based split, not a random split, when forecasting future seasons. For example, train on earlier seasons, validate on the next period, and reserve the latest seasons for final testing. If several rows belong to the same district, group-aware validation can prevent information from the same location appearing in both training and test sets.
For production work, build these controls into an implementing scalable ML pipelines for predictive analytics style workflow, with versioned data, reproducible feature generation, and logged model runs.
Build and tune the RBFN
A practical RBFN implementation has two modelling stages:
- Centre selection: Use k-means clustering on the scaled training features, or select representative observations as centres.
- Output-weight fitting: Calculate the basis-function activations and fit the output weights using ridge regression or least squares.
Start with a small grid of hyperparameters:
- Number of centres: for example, 5, 10, 20, or 40, depending on dataset size.
- Width σ: fixed from centre distances or tuned separately for each centre.
- Regularisation strength: use ridge regularisation to control unstable weights.
- Feature set: compare weather-only, historical-only, remote-sensing, and combined models.
Avoid selecting centres or scaling data using the full dataset; both operations must be fitted only on the training fold. Tune parameters using rolling-origin validation. An RBFN can overfit quickly when the number of centres is high relative to the number of district-season observations.
If you are implementing the model in Python, compare it with a custom neural-network workflow such as creating custom neural networks in Python, but keep the simpler RBFN as a transparent benchmark. For many Indian agricultural projects, reliable data pipelines matter more than adding architectural complexity.
Evaluate forecasts for decisions
Report more than one metric:
- MAE: Average absolute error in tonnes per hectare or tonnes.
- RMSE: Penalises large misses and highlights damaging forecast failures.
- R²: Indicates explained variance, but should not be used alone.
- MAPE or sMAPE: Useful only when actual values are not near zero.
- Bias: Shows whether the model consistently overpredicts or underpredicts.
Always compare the RBFN with a historical-average baseline and at least one tree-based model. Break results down by district, season, yield range, irrigation status, and forecast horizon. A model with good statewide RMSE can still fail in a drought-prone block.
Add uncertainty estimates. Bootstrap ensembles, quantile regression around the model, or conformal prediction can produce prediction intervals such as a likely lower and upper bound. Communicate these ranges clearly: a procurement team needs to know whether projected production is 100,000 tonnes with a narrow interval or a much wider risk band.
Deploy responsibly in Madhya Pradesh
A useful deployment may be a district dashboard, an API for procurement software, a mobile advisory tool, or a periodic report in Hindi and English. Show the forecast, confidence range, data date, key drivers, and known limitations. Do not present the output as a guaranteed yield or use it to deny farmer support without human review.
Monitor drift after every season. Check whether weather distributions, varieties, irrigation patterns, satellite coverage, or reporting practices have changed. Retrain only after validating the new data and preserving an immutable test period. A similar monitoring discipline is relevant to predictive analytics for Indian SME spinning mills, where operational conditions also shift over time.
Common failure modes
- Training on too few seasons and reporting unstable accuracy.
- Using post-harvest variables in an early-season forecast.
- Mixing yield and production as interchangeable targets.
- Scaling data after the train-test split incorrectly.
- Randomly splitting correlated district observations.
- Ignoring boundary changes and missing records.
- Deploying a state-level model for farm-level decisions.
- Treating correlation as agronomic causation.
The strongest project combines statistical discipline with local agronomic review. RBFNs can capture useful non-linear relationships, but their value depends on timely observations, credible labels, transparent validation, and a deployment process that farmers and institutions can actually use.