Application of hyperspectral imaging technology combined with ECA-MobileNetV3 in identifying different processing methods of Yunnan coffee beans

The quality of coffee is profoundly influenced by post-harvest processing methods. This study explored the application of hyperspectral imaging (HSI) combined with ECA-MobileNetV3 to identify five processing methods of Yunnan coffee beans. A real hyperspectral dataset was constructed and the spectra were preprocessed using Savitzky-Golay (SG) smoothing, standard normal variate (SNV), and detrending (DT). The proposed ECA-MobileNetV3 model innovatively substitutes the Squeeze-and-Excitation module with Efficient Channel Attention (ECA) mechanism, achieving enhanced feature discrimination with remarkable computational efficiency (3.36 MB model size, 4 min training time). The results demonstrated the superior performance of ECA-MobileNetV3, achieving outstanding metrics (98.40 % accuracy, 98.50 % precision, 98.35 % recall, 98.42 % F1-score), surpassing traditional machine learning models, including support vector machine (SVM), partial least squares discriminant analysis (PLS-DA), and extreme gradient boosting (XGBoost). Compared to other deep learning models, ECA-MobileNetV3 exhibits notable improvements in classification accuracy, outperforming ShuffleNetV2, EfficientNetB0, MobileNetV3 by 9.60 %, 6.00 %, 7.40 %, respectively, on the testing set. This research provides a rapid and non-destructive methodology for identifying coffee bean processing methods, offering significant potential for quality control and adulteration detection in the coffee industry.