ELMERF: A deep-learning-assisted hydroponic RGB phenotyping framework for rice seedling salt-stress evaluation and genetic mapping

Title : ELMERF: A deep-learning-assisted hydroponic RGB phenotyping framework for rice seedling salt-stress evaluation and genetic mapping

Abstract:

Rice seedling salt tolerance is commonly assessed using visual scores or survival-related indicators, but these traits are subjective, discrete, and difficult to standardize for large populations. We developed ELMERF, a deep-learning-assisted hydroponic RGB phenotyping framework for quantitative rice seedling salt-stress evaluation and genetic mapping. The framework integrates controlled hydroponic cultivation, standardized RGB imaging, RicePhenoSeg-assisted annotation and trait extraction, semantic segmentation, and genome-wide association analysis. Using this system, we constructed the Rice Seedling-Salt RGB Dataset (RSSD), in which green shoot tissue, yellow shoot tissue, roots, and background were annotated at pixel level from hydroponically grown rice seedlings under salt stress. The ELMERF network follows an encoder-decoder architecture. It combines a Mix Transformer semantic branch with an Edge-guided LoG-CNN branch and an Edge-aware Relational Fusion Head to improve the separation of slender roots, green shoots, and gradually yellowing shoot tissues. On the RSSD-A test set, ELMERF achieved a mean Intersection over Union of 51.4% and a mean Accuracy of 89.5%, outperforming nine representative semantic segmentation models. Based on the segmented shoot area, we defined shoot yellowing rate (SYR) as the proportion of yellow shoot pixels within total shoot pixels. SYR showed a smoother and more continuous phenotypic distribution than standard evaluation score and seedling death rate. In a population of 261 resequenced rice accessions, SYR was positively associated with seedling death rate and provided a quantitative description of visible salt-induced injury. Genome-wide association analysis using SYR detected 36 significant SNPs, including a major signal near the Saltol/OsHKT1;5 region, and 34 of these SNPs were not detected by conventional visual scores. These results show that standardized RGB imaging and deep-learning-based tissue segmentation can convert visible salt injury into objective and genetically informative traits. The proposed framework provides a practical route for high-throughput seedling salt-tolerance evaluation and supports the integration of digital phenotyping with rice breeding and genetic mapping.

Biography:

Dr. Jiaolong Li primarily focuses on the physiological mechanisms of stress tolerance in rice. She has made a series of advancements in the phenotypic identification of salt tolerance and the elucidation of metabolic regulatory mechanisms underlying salt tolerance in rice. Her relevant findings have been published in journals such as Plants, Agricultural Systems, and Agronomy. Some of these achievements have been granted Chinese national invention patents.