The precise online classification and identification of wheat maturity stages will offer valuable support for the intelligent control of combine harvesters. An online classification method was proposed for wheat maturity stages that combined vehicle-mounted cameras with deep learning techniques. By using real-time images captured by vehicle-mounted cameras, along with additional images from drones, a dataset of 4400 images was constructed, which included various wheat maturity stages, including milk ripening-early wax ripening stage, late wax ripening-early full ripening stage, late full ripening-dry ripening stage and harvested area. To address challenges such as complex harvesting environments and blurry wheat images, the MobileNetV2 was employed as the foundational network structure. Additionally, a convolutional block attention module（CBAM）was incorporated after feature extraction to enhance the adaptive capability of image feature extraction. To assess the credibility of the model, visualization techniques were employed to examine the areas of interest identified by the model in the images. The performance of the MobileNetV2 - CBAM model was compared with other classification models. Results indicated that the MobileNetV2 - CBAM model achieved a classification accuracy of 99.5% on the test set, which was 0.7 percentage points higher than that of MobileNetV2. When compared with ResNet and Swin Transformer models, the MobileNetV2 - CBAM model demonstrated similar classification accuracy but with a significantly smaller model memory usage（8.73 MB）—only 1/8 and1/11 of the memory usage of ResNet and Swin Transformer, respectively. Field experiments further validated the model’s practical application：at vehicle speeds of 4 km/h to 6 km/h, the system recognized an image every second with a maturity classification accuracy of 96.8%, meeting the accuracy and real-time requirements for online wheat maturity classification in mechanical harvesting scenarios.