NIRCoreVision: A novel deep learning-based framework with GUI integration for core set selection from NIRS data using 1D CNN and k-means clustering

The selection of efficient and representative core sets for nutritional quality traits is one of the major bottlenecks in plant breeding programmes. Traditional methods for generating core sets related to nutritional traits are time-consuming, labor-intensive, generates hazardous chemical waste, and require costly experiments involving significant manpower and expensive chemicals. Near-Infrared Reflectance Spectroscopy (NIRS) offers a non-destructive and rapid approach to predict the key quality traits in a high-throughput and eco-friendly manner. To address these challenges in core set selection, this study proposes a novel framework combining a 1D CNN-based autoencoder with k-means clustering to identify representative core sets from NIRS data. This approach was applied to three diverse crops (Chickpea, Cowpea, and Rice) to retain key nutritional traits content. The results showed that the core datasets achieved high representativeness with density overlap percentages ranging from 87.26% to 95.35% across specific trait content. Shannon Diversity Index (SDI) showed that over 90% of diversity was retained, such as protein in cowpea (93.6%) and chickpea (93.3%). The analysis of magnitudes at specific key wavelengths for protein and starch confirmed that the core datasets represent the complete datasets. We developed a user-friendly graphical user interface (named NIRCoreVision ) to simplify core set selection from large NIR spectral datasets. This study presents the first report of a novel and efficient high-throughput technique that integrates deep learning and clustering to develop representative core sets from large NIR spectral datasets in an eco-friendly and resource-efficient manner. • Novel 1D CNN-based framework for core set selection from NIRS data • Framework validated on chickpea, cowpea, and rice crops • Density overlaps over 85% across crops for complete and core sets • Retained over 90% Shannon diversity index (SDI) for protein and starch traits • Developed user-friendly GUI for NIRS-based core set selection