# HCA-DBN: a hill climbing optimized Deep Belief Network for crop yield classification based on kernel weight threshold

## TL;DR

This paper introduces a new machine learning model for classifying maize yield based on kernel weight, showing high accuracy even with limited data.

## Contribution

The novel HCA-DBN model combines deep learning with hill climbing optimization for improved yield classification.

## Key findings

- HCA-DBN achieved 94% accuracy in classifying maize kernel weight into low and high yield categories.
- The model outperformed standard classifiers like Random Forest and XGBoost under small sample conditions.
- Results were validated using bootstrapping and stratified 10-fold cross-validation, ensuring statistical stability.

## Abstract

Accurate classification of maize yield potential is essential for food security and effective agricultural planning, particularly in regions characterized by environmental variability and socio-economic constraints. This study explores the binary classification of maize kernel weight into low (<25 g) and high (≥25 g) categories, utilizing plant and ear traits collected from an organic maize field experiment in Vellore district, Tamil Nadu (n = 160). A Hybrid Cascade – Deep Belief Network (HCA-DBN) is proposed, utilizing the feature extraction capabilities of Deep Belief Networks (DBN) coupled with Hill Climbing Algorithm (HCA) as a lightweight hyperparameter tuning strategy. The model’s performance was benchmarked against standard classifiers including Logistic Regression, Random Forest, XGBoost, Decision Tree, Multi-Layer Perceptron (MLP), and Support Vector Classifier (SVC). The proposed HCA-DBN achieved a peak classification accuracy of 94%, demonstrating its potential to outperform conventional baselines even under small sample conditions. Rigorous validation, including bootstrapping and stratified 10-fold cross-validation, confirmed the statistical stability of the results. While these findings serve as a proof-of-concept given the dataset constraints, this study contributes a methodological benchmark for field-based maize yield classification and provides a scalable framework for future validation on larger, multi-season datasets.