Spectroscopy based novel spectral indices, PCA- and PLSR-coupled machine learning models for salinity stress phenotyping of rice.

Identification and development of salinity tolerant genotypes and varieties are one of the promising ways to improve productivity of salt-affected soils. Alternate methods to achieve this are required as the conventional methods are time-consuming and often difficult to handle large population of genotypes. In this context, hyperspectral remote sensing could be one of the rapid, repeatable and reliable methods. The aim of the present study is to develop non-invasive high-throughput phenotyping techniques for salinity stress monitoring in rice. Spectral signature of leaf samples from 56 salinity stress tolerant and sensitive rice genotypes were collected at maximum tillering and flowering stage in visible and near-infrared (VNIR) domain. The spectral reflectance data and rice leaf potassium, sodium, calcium, magnesium, iron, manganese, zinc and copper concentration were analyzed for optimum index identification and multivariate model development. Novel hyperspectral indices sensitive to leaf nutrient status as affected by salinity stress were identified. The correlation coefficient during calibration and validation of the optimized indices varied between 0.34-0.63 and 0.36-0.66, respectively. To develop multivariate model, solo partial least square regression (PLSR), PLSR- and principal component analysis (PCA)-combined machine learning models were tested. The results revealed that the performance of PLSR-combined models was the best followed by PCA-based model while indices based model was found to be least accurate. The results obtained in the present study showed potential of hyperspectral remote sensing for non-destructive phenotyping of salinity stress.