Multiscale transfer learning improves soil calcium carbonate equivalent measurement in data-limited regions using Vis-NIR spectroscopy

Abstract

Accurate measurement of soil calcium carbonate equivalent (CCE) is essential for agricultural management and carbon cycle assessments. While Vis-NIR offers a rapid and non-invasive alternative to traditional labor-intensive chemical method, its effectiveness is often constrained by regional variability and the scarcity of local datasets, limiting its applicability in data-scarce regions. Here, we introduce an innovative methodology that leverages large-scale soil spectral datasets and applies transfer learning via transfer component analysis (TCA) to enhance Vis-NIR model performance for measuring cropland soil CCE in data-scarce regions. Our TCA-based transfer models significantly outperformed locally constructed models, achieving an R² of 0.893, RMSE of 19.569, and RPD of 3.17, representing a 64.52 % improvement in accuracy. Remarkably, the proposed transfer learning strategies showcased consistent improvements even with minimal local data (R² = 0.852 and RMSE = 23.077 when only 30 local samples were available), highlighting their robustness and scalability. These findings demonstrate that the integration of transfer learning with Vis-NIR offers a reliable solution for soil CCE measurement in regions lacking sufficient local data, advancing the broader application of spectral analysis in soil science and contributing to more effective agricultural practices.