The key contributions of this work are highlighted as follows:
a) The NirMACNet architecture, which leverages deep learning, is proposed for predicting milk composition. The model utilizes multi-scale convolutional kernels to extract features from raw NIR spectra across different scales. This multi-scale feature extraction enhances the model's robustness and ability to represent the captured characteristics, even under non-ideal measurement conditions. Furthermore, the architecture incorporates residual connections to facilitate learning in profound networks, specifically addressing the performance degradation issue observed in traditional deep learning models for NIR spectra analysis.

b) The NirMACNet architecture capitalizes on the strengths of CNNs for feature extraction and regression directly from raw NIR spectra data. This eliminates the need for complex preprocessing steps, streamlining the overall analysis process and potentially reducing the risk of information loss during preprocessing.