Cross-Site Prediction of Soil Organic Carbon in Plantation Forests Using Vis-NIR Spectroscopy and Target-Spectrum-Guided Source-Domain Weighting

Abstract

Soil organic carbon (SOC) is an important indicator of plantation-forest soil quality, but Vis-NIR models calibrated at one site may show systematic bias at another. This study used 370 Gaofeng samples as the source domain and 119 Yachang samples as the target domain to evaluate SOC-gradient-constrained spectral similarity weighting (SOC-SSW). Unlabeled target spectra guided source-sample weighting, while source SOC strata maintained calibration-gradient coverage. In Gaofeng-to-Yachang transfer, direct PLSR produced R2 = 0.809, RMSE = 6.676 g kg−1, and bias = −3.842 g kg−1. SOC-SSW increased R2 to 0.855, reduced RMSE to 5.800 g kg−1, and decreased absolute bias by 75.42%. Paired bootstrap analysis confirmed lower RMSE and MAE, and paired absolute errors remained significantly lower after Holm correction (p = 0.0018). Similar accuracy was retained when weights were constructed from 10 or 40 target spectra, although the complete target set yielded the smallest absolute bias. Reverse transfer was unsuccessful, indicating direction-dependent applicability. SOC-SSW can therefore support conditional reuse of laboratory soil spectral datasets when target-site SOC labels are unavailable, provided that the source dataset is sufficiently representative.