CHENNAI: Researchers from Tamil Nadu have developed a crop prediction model that delivers nearly 98 per cent accuracy by factoring in the interrelationship between soil nutrients and climatic conditions, a gap they say has long weakened conventional systems.
Chennai-based scientists from the School of Computer Science and Engineering at VIT and collaborators from Kanniyakumari have proposed a new classification framework designed to improve crop recommendation decisions using structured feature correlation analysis. Their findings have been published in the journal Scientific Reports.
Crop selection remains a critical decision for farmers, particularly in regions facing erratic rainfall, soil variability and temperature fluctuations. While existing prediction systems rely on standard classification techniques, the research team found that most models assess agricultural variables independently, overlooking how factors such as nitrogen, phosphorus, potassium, rainfall and humidity influence one another.
"Our study addresses a key limitation in current crop prediction systems, the lack of correlation analysis between features.
By extracting and integrating feature relationships before classification, the prediction performance improves substantially, " the authors said in the paper. The researchers introduced what they call a Feature Correlation Square-based Nearest Neighbor (FCSNN) approach. The system first standardises agricultural inputs such as soil nutrients, pH levels, rainfall, humidity and temperature.
It then constructs a correlation matrix using statistical measures to identify how these variables interact. The refined dataset is subsequently used to classify the most suitable crop.
The model was tested using a publicly available crop recommendation dataset comprising 2,200 records across 22 crops, including rice, maize, cotton, coconut, banana, coffee and pulses. Each crop category contained 100 records, ensuring balanced evaluation conditions.
According to the study, the proposed system achieved 97.9 per cent accuracy, outperforming widely used classifiers such as Decision Tree, Naive Bayes, Logistic Regression, Random Forest, XGBoost and conventional k-Nearest Neighbor models. The error rate was reported at 2.1 per cent.
The authors noted that selecting the optimal number of nearest neighbours was crucial. "Empirical testing showed that a value of four neighbours produced the most consistent results," the paper stated.
Agriculture experts say the framework has practical relevance, particularly for states like Tamil Nadu where farmers must make crop decisions amid climate variability and resource constraints. The model can assist decision-makers at district and state levels in aligning crop planning with prevailing soil and weather conditions.
The researchers indicated that future work could involve real-time field data collection using sensor-based systems to further refine crop recommendations.