A Comparative Study of Recurrent Architectures for Imbalanced Dry Bean Classification

Abstract The classification of dry bean varieties is vital for agricultural productivity and food quality control. However, the inherent class imbalance in datasets poses challenges for machine learning (ML) and deep learning (DL) models. This study evaluates the performance of three DL models-Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and One-Dimensional Convolutional LSTM (ConvLSTM1D)-in classifying 13,611 samples of seven dry bean varieties from the publicly available Dry Bean Dataset. The ConvLSTM1D model’s novelty lies in its hybrid architecture, which integrates 1D convolutional layers to extract complex local patterns from the morphological features before recurrent processing. This structural advantage allows it to create more discriminative feature representations, effectively addressing the class imbalance by better identifying minority classes without explicit resampling techniques. Experimental results show that ConvLSTM1D achieves the highest performance, with an accuracy of 93.78% and macro-averaged precision, recall, and F1-scores of 93.65%, 93.78%, and 93.70%, respectively. This model also demonstrates a significant reduction in misclassifications and maintains robust performance despite dataset imbalance. In contrast, LSTM and GRU models exhibit comparable but slightly lower classification accuracy and are less effective in differentiating visually similar bean types. This research highlights the potential of ConvLSTM1D as a reliable and accurate method for automated dry bean classification, contributing to improved efficiency and precision in agricultural practices.