Soil total nitrogen inversion and interpretability analysis using vis-NIR spectroscopy and transfer learning

The use of Vis-NIR Spectroscopy for soil component inversion has increased, driven by its advantages in non-destructive, large-scale monitoring. However, it often faces challenges in model generalization. Transfer learning, leveraging large existing soil sample datasets, is considered an effective solution to overcome the mentioned limitations. This study explores the use of transfer learning, utilizing the LUCAS database, to improve the accuracy of estimating soil total nitrogen (STN) with Vis-NIR spectroscopy in Gejiu, Yunnan, China, addressing challenges in model generalization. It compares spectral preprocessing methods (logR, SNV, MSC) and models (PLS, RF, ResNet) to assess their impact on inversion performance. SHapley Additive exPlanations (SHAP) is employed for model interpretability. Results show that transfer learning with the ResNet model significantly enhances STN inversion, particularly with MSC preprocessing, where the average R2 improves from 0.51 to 0.70. Among the models tested, ResNet with transfer learning outperforms others in accuracy. SHAP analysis identifies key wavelengths -2050, 2459, 2149, 2109, 2410, and 1470 nm - as crucial for predicting STN, which aligns with the observed correlations. This research validates the effectiveness of transfer learning on small datasets, offering a robust solution for STN inversion using Vis-NIR spectroscopy.