Accurate and timely crop yield estimation is of great significance for ensuring food security and promoting sustainable agricultural development. Focusing on the Guanzhong Plain in Shaanxi Province, selecting the vegetation temperature condition index (VTCI), leaf area index (LAI), and fraction of photosynthetically active radiation (FPAR) as remotely sensed feature parameters closely related to winter wheat growth. By combining the advantages of long short-term memory network (LSTM) in processing time series data with the capability of the attention mechanism (AM) to extract key information, an LSTM-AM deep learning model for winter wheat yield estimation was constructed. The results indicated that the coefficient of determination (R2) for the LSTM-AM model was 0.65, with a root mean squared error (RMSE) of 496.43 kg/hm2 and a mean absolute percentage error (MAPE) of 7.31%, and a normalized root mean squared error (NRMSE) of 0.11. Compared with the LSTM model (R2=0.60, RMSE is 527.81 kg/hm2, MAPE is 7.85%, NRMSE is 0.15), the LSTM-AM model demonstrated higher estimation accuracy and mitigated the underestimation of high yields and overestimation of low yields observed in the LSTM model. To further enhance the model's interpretability, the permutation feature importance (PFI) method for an in-depth explanation of the estimation results was utilized. The findings indicated that FPAR from late April to early May, VTCI from late March to early May, and LAI from late April to late May significantly influenced the final yield of winter wheat. Therefore, the LSTM-AM model not only exhibited high estimation accuracy but also enhanced interpretability through the PFI method, providing a valuable reference for subsequent winter wheat yield estimation.