Ensemble-Based Deep Learning Framework for Automated Wheat Variety Identification Using Digital Seed Morphometry

Abstract

Correct wheat variety identification maintains seed purity and supports breeding programs. It also improves precision agriculture through reliable seed classification. Traditional visual identification needs human effort and expert judgement. These methods work slowly and often produce errors. Problem increases when different wheat varieties show similar physical features. This study proposes an ensemble-based deep learning framework for automatic wheat variety identification. System uses digital seed images and morphological features of wheat seeds. Method includes image acquisition, preprocessing, feature extraction, model training and ensemble fusion.  Study integrates all steps into one classification pipeline. It classifies twelve Indian wheat varieties using extracted seed images.  Framework extracts geometric and surface features from segmented seed images. Study developed classifiers based on ANN, CNN, SVM, RF, k-NN and LR. It combines model outputs through F1-score weighted soft voting. It also uses logistic regression as a meta-learner for stacking. Study evaluated the model through stratified five-fold cross-validation. Proposed ensemble model achieved 92.8% overall accuracy. It achieved 0.92 macro precision and 0.93 macro recall. It also achieved a macro F1-score of 0.92. One wheat variety achieved 100% correct classification. All remaining varieties achieved classification accuracy above 81%. System works as a non-destructive, scalable and computationally efficient tool. It supports seed certification, breeding programs and agricultural quality control. Results show that the hybrid ensemble algorithm classifies twelve wheat varieties accurately. It performs with strong stability and efficiency across different varieties. It outperforms individual models and provides a reliable tool for seed certification.