Abstract
Infrared thermal imaging offers a rapid and sensitive approach to assessing temperature changes in plants caused by salt stress, even in the early stages of exposure. Given the increasing prevalence of salt contamination in the environment, it is essential to accurately estimate salinity levels, as the effects strongly depend on salt concentration: moderate salinity elicits a reversible, osmotic driven rise in leaf temperature, whereas higher salinity induces a larger, sustained temperature increase indicative of ion toxicity related stress. We propose a method to evaluate the severity of salt stress in plants exposed to sodium chloride, using a series of thermograms captured through a non-invasive infrared imaging technique under illuminated conditions. Thermal measurements are then used to train machine learning models used to perform multi-class classification to distinguish between four different salt concentrations. To test the proposed method, we cultivated Arabidopsis thaliana plants under controlled conditions. Data collected from the prepared samples were used to assess the accuracy of various approaches and classifiers with lead-one-out cross-validation. This experimental evaluation shows that the optimal performance is achieved when the datasets used for training consist of longer sequences of thermal data provided to models using neural networks.