Understanding plant vulnerability to stressors with automated image analysis

Plants often experience adverse or stressful environments that might have an impact on their growth and development, thus, phenotype. Visible symptoms of stress had been long studied, but their manual scrutiny can be challenging, time-consuming, and error-prone. However, there are currently very few instances of machine learning (ML) models that can automatically predict plant stresses, especially abiotic ones, from image-derived morphological traits. This study aims to fill this gap using digital phenotyping tools for stress detection based on automated image analysis and to further compare them with standard analytical procedures commonly carried out in laboratories. Two preliminary models for classifying salt stress levels have been developed to achieve this goal on Arabidopsis thaliana plants. Seedlings were grown on different substrates (soil and perlite) and exposed to “medium” and “high” salinity stress levels (50 mM and 150 mM NaCl) for 10 days. Biochemical parameters – Electrolyte Leakage (EL), Relative Water Content (RWC), and Dry weight (DW) – were measured, along with morphological traits – colorimetrical and geometrical – obtained from RGB images using both manual and automated approaches. The resulting data was then used to evaluate the performance of decision trees on 2-classes (presence or absence of stress) and 3-classes models (absence, medium, and high-stress levels). We noticed that plants’ development was influenced both by the growing environment and substrate type. Visible symptoms of stress included a reduction in leaf number and rosette size, as opposed to a chlorosis increment. RWC and DW decreased in response to high NaCl concentration, whereas EL increased. Nevertheless, while differences were significant among high-stressed plants and control ones, medium-stressed plants were hard to discern from both conditions. The Principal Component Analysis, which grouped the two levels of stress, also supported this conclusion. These results were further validated by classification algorithms tested: the 3-classes model only achieved 73% accuracy, compared to the binary model’s 90%. EL appears to be one of the key features for stress detection, but other important image-derived functional traits have also emerged from this preliminary study that can be used as indicators of plant health status and to study plant strategy to cope with environmental stressors, thus predicting their vulnerability/resilience to extreme climate conditions. With these findings, we demonstrate the great potential of image analysis methods and we highlight the positive impacts of automation, including increased analysis speed and decreased error rates. Moreover, we emphasize the importance of explainable ML models that can be easily interpreted and indicate essential traits needed for the model’s deployment and improvement. Given the modest size of our dataset, an integrated strategy is still necessary to obtain an adequate degree of classification accuracy. Therefore, in the future, by increasing the number of instances, we aim to enhance the model’s robustness and reliability and generalize it for the detection of various abiotic stresses in diverse species.