Hyperspectral Sensing for High-Throughput Chloride Detection in Grapevines

Abstract

Soil salinity affects major viticultural areas worldwide with chloride ions being the primary source of salt toxicity in grapevines. This toxicity impacts vine health and reduces fruit yield and quality. Current breeding efforts to improve grapevine salinity tolerance are limited by the low throughput of available phenotyping methods, which are time-consuming, labor-intensive, and destructive. This study demonstrated that hyperspectral proximal sensing can be utilized as a high-throughput, non-destructive screening technique to identify salinity-tolerant grapevine germplasm. The predictive abilities of two different hyperspectral devices, which varied in price, resolution, and sensitivity, were compared across 23Vitisaccessions spanning eight species. Prediction models were built using hyperspectral reflectance and leaf chloride content measured with a lab chloridometer. Three distinct approaches were studied: 1) analyzing the correlation between individual wavelengths and chloride content; 2) employing machine learning models, including Partial Least Squares Regression (PLSR), Random Forest (RF), and Support Vector Machine (SVM), utilizing all wavelengths; and 3) classification-based prediction using Partial Least Squares Discriminant Analysis (PLSDA). Multiple regions in the spectrum, including 613-660 nm, 689-696 nm, and 1357-1358 nm, showed a medium correlation (0.30-0.50) with chloride content in the leaves. PLSR was the most effective machine learning approach, demonstrating moderate predictive capability for chloride content (maximumR²= 0.67), though performance varied between the two devices tested. With PLSDA, predictions increased considerably, up to an accuracy of 0.97, depending on the instrument used and the spectral data transformation. Overall, the more expensive and sensitive device with a wider spectral range outperformed the more affordable, shorter-range device. However, when the prediction model was based on classes (chloride excluders vs. non-excluders) rather than chloride content, the differences in prediction abilities were minimal, with both instruments performing very well. This is promising for identifying breeding materials with chloride exclusion capabilities at low cost and high throughput.