Advanced machine learning techniques for hyacinth bean identification using infrared spectroscopy and computer vision†

Abstract

The classification and quality assessment of underutilized hyacinth bean (HB) (Lablab purpureus L.) landrace accessions were systematically performed using state-of-the-art machine learning (ML) approaches. Invasive and non-invasive techniques were used to identify and evaluate the accessions via FTIR and computer vision, respectively. Regression and classification models based on FTIR achieved outstanding accuracy in chemical characterization; among these, neural network models demonstrated better performance in terms of R², RMSE, and computational efficiency. However, sample preparation and scalability posed challenges for high-throughput applications. The non-invasive techniques fared better when a transfer learning approach was applied using the pretrained model EfficientNet_V2_S, achieving an F₁ score of 98.25% for classification. These methods could also offer lower computational costs and minimal preprocessing. Comparative investigations revealed the advantages of each approach: the accuracy of chemical analysis through the FTIR technique and the scalability/resource efficiency of computer vision. The predictive accuracy was further improved in the neural network model and KNN technique employing hyperparameter tuning, highlighting the need for systematic tuning techniques. This paper highlights the need for hybrid methods that combine invasive and non-invasive strategies for the comprehensive identification of HB accessions. This study presents practical methodologies for classification and quality assessment that support sustainable agricultural practices, enhance biodiversity conservation efforts, and optimize crop management strategies while facilitating the integration of advanced ML technologies into agriculture and food research.