The identification and selection of drought-tolerant wheat varieties are of great significance for ensuring food security and sustainable agricultural development in China.The root system serves as the primary pathway for water absorption in plants, and the phenotype of the root system is closely related to drought tolerance. In order to obtain root system phenotypic indicators of wheat quickly and accurately,drought stress experiments were conducted by using the soil box method and collected sequential root images at 18 time points. A deep learning-based image processing and analysis workflow was proposed for root image processing. To address the issue of root breakage caused by soil occlusion, a two-stage detection-repair method combining object detection networks and hourglass attention networks was designed to repair the broken root regions. Multi-scale training and adaptive iteration were employed to improve the accuracy and robustness of the repair process. Six phenotypic traits of wheat root systems, including root area, total root length, and root width, were extracted under drought stress and control conditions, and the phenotypic responses of wheat roots to drought stress were analyzed. The results showed that under drought stress, wheat exhibited lower root biomass, deeper rooting depth, and more dispersed root architecture. The drought tolerance index of wheat root systems was calculated, and the drought tolerance of wheat varieties was described and ranked by using principal component analysis. A method of in-situ measurement and analysis of wheat root phenotype was proposed, which can be applied to the study of wheat drought resistance.