To address the challenge of agricultural greenhouse temperature measurement being highly susceptible to noise, which limits direct prediction accuracy, this study proposes an integrated prediction model, XGB-KF, combining XGBoost and the Kalman filter. First, the model estimates the current greenhouse temperature using XGBoost. Then, the Kalman filter dynamically adjusts the estimated result to refine the prediction. Numerical experiments are conducted using sensor data from a greenhouse in Zhuozhou, with root mean square error (RMSE) as the main evaluation metric. Compared with XGBoost, Bi-LSTM, and Bi-LSTM-KF methods, the XGB-KF model reduces RMSE by 5.22%, 10.85% and 7.45% respectively.