Assessment of salt tolerance in peas using machine learning and multi-sensor data

• Estimation plant height and canopy coverage of pea under salt stress and control conditions based on RGB images. • Estimation of pea biomass and SPAD values under two growth conditions (control and saline-alkali) based on multi-source remote sensing data. • To construct a comprehensive salt tolerance score for peas using the estimated data and ground-collected data, enabling high-throughput evaluation of pea salt tolerance. Salt-alkali region spans vast areas and holds significant potential for agricultural development. Screening for salt-tolerant crop varieties is a critical strategy to enhance and utilize such region. Among edible legume crops, the peas are notable for their short growing period and moderate salt tolerance, making them a promising candidate for cultivation in salt-alkali conditions. Accurate and efficient screening of salt-tolerant pea varieties is essential for improving these regions. However, traditional screening methods are often time-consuming, labor-intensive, and prone to human error. Recent advancements in Unmanned aerial vehicle (UAV) and sensor technologies have enabled high-throughput screening of salt-tolerant crops, offering a more efficient alternative. In this study, UAVs equipped with red-green-blue (RGB) and multispectral (MS) sensors were deployed to capture images of peas grown in both normal and salt-treated plots. Structural traits (PH and canopy coverage [CC]), texture features, and spectral data were extracted from these images. Using this information, aboveground biomass (AGB) and Soil Plant Analyses Development (SPAD) values were estimated under both growth conditions using four machine learning algorithms: CatBoost, Light Gradient Boosting Machine (LightGBM), support vector machines (SVM), and random forest regression (RF). To asses salt tolerance, pea salt tolerance score (PSTS) was developed based on four indicators—plant height (PH), CC, AGB, and SPAD values. The score was then compared with ground-based measurements to validate its accuracy. The results show that: 1) multi-source data fusion significantly improved the accuracy of AGB and SPAD estimation; 2) the CatBoost algorithm achieved the highest performance for AGB estimation (R² = 0.70, RMSE = 1.59 t/hm 2 , NRMSE = 13.94%), while the LightGBM algorithm performed best for SPAD estimation (R² = 0.60, RMSE = 2.33, NRMSE = 14.53%); and 3) The PSTS established based on the optimal estimation data exhibits a strong consistency with the ground-measured data. In conclusion, integrating multi-sensor data and with advanced machine learning techniques provides a feasible and reliable approach for screening salt-tolerant pea varieties, paving the way for better utilization of salt-alkali region.