Abstract: To achieve lossless and efficient monitoring of physicochemical changes in tobacco leaves during the initial curing process, a lightweight MobileNetV2_CBAM deep transfer learning regression model is proposed. This model employs the pre-trained MobileNetV2 as its feature extraction backbone network, incorporating a convolutional attention module (CBAM) and a deep regression head at the end to enhance feature expression capability and nonlinear fitting ability. Using the MobileNetV2_CBAM model to predict moisture content, starch content, reducing sugar content, and total sugar content of tobacco leaves during initial curing, the results indicate: 1) Simultaneously incorporating CBAM and a deep regression head into the MobileNetV2 model significantly improves predictive performance. 2) The MobileNetV2_CBAM model achieves 0.561 G floating-point operations (FLOPs) with total parameters of 15.43 M. The coefficient of determination (R²) for predicting the four physicochemical indicators ranged from 0.8651 to 0.9065, the mean absolute error (MAE) ranged from 0.0622 to 0.9831, the root mean square error (RMSE) ranged from 0.0927 to 2.0435, Residual Prediction Deviation (RPD) ranged from 2.5943 to 4.6462, and Frames Per Second (FPS) ranged from 100.5780 to 121.2540. This model achieved a favorable balance between predictive performance, processing speed, and deployment cost. 3) Across different baking stages, the MobileNetV2_CBAM model achieved an explanatory variance score ranging from 0.73 to 0.91 for predicting four physicochemical indicators, demonstrating strong predictive capability. 4) In complex backgrounds, the R² values for the MobileNetV2_CBAM model's predictions of the four physicochemical indicators ranged from 0.7885 to 0.8701, maintaining strong predictive capability. This study provides theoretical reference for online monitoring of physicochemical indicators during the initial curing process of tobacco leaves.