Abstract: Accurate estimation of winter wheat yield under abnormal climate conditions is essential for understanding the impacts of climate change on crop production and ensuring grain security. Focusing on 286 counties in the major winter wheat production regions of China, this paper constructs an 8-day ¹county-level yield estimation dataset based on multi-source remote sensing vegetation indices and meteorological variables from the green-up to maturity stages during 2002-2020. Furthermore, applying machine learning and deep learning approaches, this paper systematically evaluates the estimation performances of different models, including the convolutional neural networks (CNN), the long short-term memory networks (LSTM), the XGBoost, and the random forest (RF), under drought, flood, and low-temperature stress conditions. Results show that transfer learning significantly improves model accuracy and stability, with overall RMSE reductions of 18.1%~19.6% for deep learning models. Among all models, the transfer-learning-based CNN demonstrates the most robust overall performance, achieving an R² of 0.63 and an RMSE of 722.86 kg/hm² under five-fold cross-validation; under specific disaster scenarios, the CNN performs best in drought and low-temperature years, while the LSTM outperforms other models in flood years. Spatial analysis indicates that 55%~64% of counties have yield estimation errors within 10%, with generally lower errors in the central and eastern plains and relatively higher errors in western and southwestern hilly regions. The proposed CNN model integrating multi-source data and transfer learning demonstrates strong robustness and generalization capability under anomalous climate conditions, providing an efficient regional-scale approach for winter wheat yield estimation to support agricultural management and food security decision-making.