Abstract
Near-Infrared Spectroscopy (NIRS) analysis technology faces numerous challenges in industrial applications. Firstly, the generalization capability of models is significantly affected by instrumental heterogeneity, environmental interference, and sample diversity. Traditional modeling methods exhibit certain limitations in handling these factors, making it difficult to achieve effective adaptation across different scenarios. Specifically, data distribution shifts and mismatches in multi-scale features hinder the transferability of models across different crop varieties or instruments from different manufacturers. As a result, the large amount of previously accumulated NIRS and reference data cannot be effectively utilized in modeling for new instruments or new varieties, thereby limiting improvements in modeling efficiency and prediction accuracy. To address these limitations, this study proposes a novel transfer learning framework integrating multi-scale network architecture with Balanced Distribution Adaptation (BDA) to enhance cross-instrument compatibility. The key contributions include: (1) RX-Inception multi-scale structure: Combines Xception's depthwise separable convolution with ResNet's residual connections to strengthen global-local feature coupling. (2) Squeeze-and-Excitation (SE) attention: Dynamically recalibrates spectral band weights to enhance discriminative feature representation. (3) Systematic evaluation of six transfer strategies: Comparative analysis of their impacts on model adaptation performance. Experimental results on open corn and pharmaceutical datasets demonstrate that BDSER-InceptionNet achieves state-of-the-art performance on primary instruments. Notably, the proposed Method 6 successfully enables NIRS model sharing from primary to secondary instruments, effectively mitigating spectral discrepancies and significantly improving transfer efficacy.