A practical UAV-assisted workflow for high-throughput screening of drought-resistant rice

Rice is the main food for half of the global population, but drought is one of the main abiotic stresses affecting rice yield. Therefore, breeding rice varieties with enhanced drought resistance is an essential strategy to ensure food security. Varieties with stronger drought resistance typically exhibit lower degrees of leaf withering and higher yields under drought. However, no automatic screening model for drought resistance in rice has yet been established that accounts for leaf wilting and post-drought yield. To improve the screening efficiency and accuracy, it is urgent to develop a standardized pipeline for high-throughput drought-resistance rice evaluation. This study established a set of standardized workflows for screening drought-resistance rice in complex population based on unmanned aerial vehicle (UAV) images. Firstly, we defined five levels (Level 1–5) of rice leaf withering and developed an efficient and automatic UAV RGB (red-green-blue) image processing software PLOT ASSISTANT. Secondly, based on manual measurement and UAV images, we constructed a leaf withering assessment model and a rice drought-resistance prediction model in terms of final yield under drought stress. Among leaf withering assessment models, the MobileNet model introducing spatial attention mechanism based on transfer learning exhibited the highest assessing accuracy (R2 =0.94, MSE=0.08). The rice drought-resistance prediction model introducing coordinate inverse projection successfully screened the top 5 % of drought-resistant rice varieties with the highest yield (F1 = 0.95, Precision=0.95, and Recall= 0.96). Using our method, the screening of the desirable drought-resistant rice varieties could be accomplished at the early growth stage, and will drastically reduce time and labor consumption for manual trait measurement during traditional screening. Furthermore, the screening model could work in complicated big population including not only Japonica rice but also Indica or median type between Japonica and Indica. Last, using the image-traits from the UAV, we also identified 1942 novel genetic loci associated with drought-resistance providing valuable resources for future research and breeding aimed at enhancing drought resistance of rice. • Large-scale UAV rice drought dataset of 2979 plots across two years. • Automated UAV image preprocessing software "PLOT ASSISTANT". • First yield-based model predicts rice drought resistance. • High-throughput screening pipeline for accelerated breeding selection. • Novel drought-resistance QTLs revealing genetic mechanisms.