Abstract: Air environmental information plays an important role during plant growth and reproduction, and prompt and accurate prediction of atmospheric environmental data is helpful for agricultural robots to make timely decisions. In the interest of efficiency, an online learning method for predicting air environmental information was presented in this work. This method combines the advantages of convolutional neural network (CNN) and experience replay technique: CNN is used to extract features from raw data and predict atmospheric environmental information, while experience replay technique can store environmental data over some time and update the hyperparameters of CNN. To validate the effects of this method, this online method was compared with three different predictive methods (including random forest, multi-layer perceptron, and support vector regression) using a public dataset (Jena). According to the results, a suitable sample sequence size (e.g., 16) has a smaller number of training sessions and stable results; a larger replay memory size (e.g., 200) can provide enough samples to capture useful features; and 6 d of historical information is the best setting for training predictor. Compared with traditional methods, the method proposed in this study is the only method that can be applied for various conditions.