LMS-Net: A Light-Weight Network for Mungbean Salt Stress Identification

Abstract

Salt stress is one of the main factors affecting mungbean production, and the assessment of salt tolerance in mungbean typically requires long-term morphological observation or time-consuming physiological and biochemical experiments, which are time-consuming and labor-intensive in field conditions. This study introduced a salt stress level identification model for mungbeans, named LMS-Net, which used early-stage chlorophyll fluorescence imaging and a lightweight convolutional neural network (CNN) model to identify the salt tolerance of seedlings. To facilitate deployment on mobile and portable devices, we incorporated depth-wise separable convolutions and reparameterization techniques into LMS-Net to reduce latency during inference. Moreover, to better capture the key features and regions of stress, we proposed a Multi-scale Collaborative Attention Mechanism (MSCA) into LMS-Net. Using the FluorCam Open Chlorophyll Fluorescence Imaging System, we collected 1,879 chlorophyll fluorescence images of mungbean natural populations under salt stress to create a mungbean salt stress chlorophyll fluorescence dataset. Through comprehensive comparisons with mainstream lightweight models based on CNN and Transformer architectures, the LMS-Net model achieved an average identification accuracy of 97.64%, an average precision of 97.78%, an average recall of 97.74%, and an average F1-score of 97.75% on the test set, which benefit from model architecture and MSCA. These classification performance metrics surpass those of the comparative models. Using this method, we identified two highly salt-tolerant mungbean materials, ‘Suhuang 1’ and ‘Wei 9002-341’, from a natural population of 230 mungbean accessions. In summary, this method is promising and valuable for early-stage salt tolerance assessment of mungbean seedlings, offering new insights for screening salt-tolerant germplasm resources in mungbeans and other crops.

Keywords

• Mungbean salt stress
• salt tolerance
• chlorophyll fluorescence imaging
• lightweight CNN
• structural reparameterization
• attention mechanism