Towards non-destructive monitoring of plant salinity stress: an integrated approach combining image analysis and machine learning

Giorgia Del Cioppo ¹, Simone Scalabrino ², Gabriella Stefania Scippa ¹, Dalila Trupiano ¹

University of Molise ¹, University of Molise; Datasound srl ²

Abiotic stresses, such as salinity, significantly impact plant growth, necessitating the development of effective and scalable stress detection tools. In this context, investigating digital traits from image-based phenotyping could be crucial for identifying rapid and non-destructive stress markers. This study explores the feasibility of integrating image-based phenotyping with minimal biochemical analyses to classify salt stress using machine learning. Two plant species were selected: Arabidopsis thaliana (AT, var. Col-0) as a model organism, and Phaseolus vulgaris (var. Fagiolo d'acqua - FDA, and Fagiolo della levatrice - FDL) as a non-model species. Stress responses were evaluated under control, medium, and high-salinity treatments. The experimental workflow included three steps: feature extraction from image and laboratory data (i), data preprocessing for cleaning and feature selection (ii), and the training, testing and evaluation of classification models (iii). ImageJ software and the ARADEEPOPSIS pipeline (Hüther et al., 2020) were used for image segmentation and digital feature extraction, which included geometric and colorimetric parameters. Laboratory analyses focused on 16 traits, with malondialdehyde (MDA) content and electrolyte leakage (EL) emerging as particularly sensitive indicators of stress severity. The two species exhibited distinct responses to salt stress, with Phaseolus vulgaris populations showing greater variability in pigment contents and morphological traits compared to Arabidopsis thaliana. Specifically, AT plants displayed clear pigment degradation, evident in increased chlorosis along the stress gradient, reflecting a stress-induced reduction in chlorophyll. Conversely, FDA and FDL demonstrated an increase in carotenoid levels, likely functioning as antioxidants. Principal Component Analysis (PCA) and Correlation Analysis (CA) identified significant relationships between laboratory traits and image-derived ones: biomass, measured as both fresh weight (FW) and dry weight (DW), was correlated with various geometric traits, such as area, perimeter, and equivalent diameter. Additionally, malondialdehyde (MDA) content and electrolyte leakage (EL) were correlated with Chroma Difference, while pigment contents were correlated with Chroma Ratio and Green Strength. Decision trees and random forests algorithms were trained using selected features to obtain a 2-class model, to identify stress presence, and a 3-class model, to classify stress intensity. Automated image-derived traits alone demonstrated a precision of 0.88 for binary classification and 0.77 for multi-class classification respectively, highlighting their robustness as proxies for stress detection. However, integration with selected laboratory-derived features enhanced model accuracy and reliability. The best results were indeed obtained using combined features, achieving 0.91 precision in distinguishing stressed from unstressed plants and a 0.84 precision in classifying stress intensity. The study successfully demonstrated that a combination of image analysis and reduced biochemical analysis can effectively differentiate stressed and unstressed plants, showing the potential of digital features to reflect physiological changes and act as non-destructive stress proxies. This accessible approach not only ensured reduced computational demands but also highlighted the importance of integrating image-derived traits with few traditional biochemical markers to enhance model accuracy. Future work will address the need for larger datasets to further validate these findings and improve the generalizability of machine learning models, as well as expanding this framework to additional abiotic stressors (e.g., drought, heat).

Main author career stage: PhD student

Contribution type: Poster

First choice session: 4. Structure, physiology, and development

Second choice session: 3. Biodiversity and global change