توصيفگر ها :
مدلسازي نرم و مستقل شباهتهاي بين كلاسي (SIMCA) , تفكيك حداقل مربعات جزئي (PLS-DA) , ماشينهاي بردار پشتيبان (SVM) , شبكههاي عصبي مصنوعي (ANN) , انتخاب ويژگي , ، رگرسيون فرآيند گوسي (GPR) , جنگل تصادفي (RF) , درخت تقويت شده (BT)
چكيده انگليسي :
Given the challenges of climate change, global warming, and increasing drought, reducing water consumption and improving water use efficiency in agriculture, especially in arid regions, has become critically important. In these regions, drought stress is one of the primary factors limiting crop growth and productivity. Early detection of drought stress is essential to prevent irreversible damage to plants and minimize yield loss. In this study, hyperspectral imaging at the leaf level and thermal imaging at the canopy level were utilized to assess drought stress in safflower plants. Three safflower genotypes, Palenus, A82, and IL-111, were cultivated under three irrigation treatments corresponding to 50%, 70%, and 90% soil water content depletion. These treatments represented unstressed (US), mild stress (MS), and severe stress (SS) conditions, respectively. Hyperspectral images of the leaves were captured before visible signs of water deficiency appeared. Stress classification was conducted using the full average spectral data with modeling techniques such as soft independent modeling of class analogy (SIMCA), partial least squares discriminant analysis (PLS-DA), support vector machines (SVM), and artificial neural networks (ANN). Feature selection methods, including SIMCA-based selection, PLS-VIP, and CARS, were employed to identify effective wavelengths. ANN models were then used to classify stress levels based on the selected wavelengths. Spatial analysis was conducted using pixel-level classification through unsupervised (k-means clustering) and supervised (the best-performing classifier model) approaches. Additionally, mean leaf spectra were applied to develop regression models for estimating relative water content (RWC). Regression techniques such as partial least squares regression (PLSR), Gaussian process regression (GPR), ANN, random forest (RF), and boosted trees (BT) were implemented. Thermal images collected under field conditions were analyzed to extract leaf temperature histograms, from which features such as the maximum, minimum, range, mean, median, mode, skewness, and kurtosis of canopy temperature were derived. Crop Water Stress Index (CWSI) and stomatal conductance index (Ig) were also calculated from the thermal image grayscale values, and their relationships with RWC and measured stomatal conductance (gs) were examined. The results revealed that ANN models using full spectral data outperformed other classifiers in distinguishing US, MS, and SS classes, achieving weighted F1-scores of 92.22%, 96.01%, and 96.47% for the Palenus, A82, and IL-111 genotypes, respectively. Among feature selection methods, SIMCA-based selection performed best in monitoring stress conditions for Palenus and A82. In supervised spatial analysis, ANN models effectively depicted the progression of stress in the leaves of different genotypes. RWC prediction models demonstrated that RF with second derivative preprocessing, with MSC preprocessing, and with second derivative preprocessing achieved the best predictive accuracy for the Palenus, A82, and IL-111 genotypes, respectively. The RPD values of 2.035, 2.268, and 2.263 for Palenus, A82, and IL-111 indicated good predictive performance. Thermal image analysis showed negative correlations between RWC and CWSI (R² = 0.358) and between gs and CWSI (R² = 0.328), as well as positive correlations between RWC and Ig (R² = 0.361) and gs and Ig (R² = 0.332). This study underscores the potential of hyperspectral imaging for distinguishing drought stress levels and estimating RWC in different safflower genotypes, highlighting its utility for managing this important oilseed crop.
