Selecting the right samples rather than more samples: A new spectral–environmental similarity strategy for local soil spectral modeling

Addressing the dual challenges of limited sample size and high environmental heterogeneity in small-scale soil organic carbon (SOC) spectral modeling, this study proposes a fundamental hypothesis: selecting samples that are similar to the target region in both “spectral features and environmental characteristics” is more effective for improving prediction accuracy and stability. Based on this assumption, we developed a synergistic sample transfer strategy that integrates spectral similarity with environmental similarity under the Third Law of Geography, aiming to systematically screen the most comparable samples from the global soil spectral library to enhance the performance and robustness of local SOC modeling. A spectral-environmental similarity framework was established to identify samples that are simultaneously similar to the target region in spectral properties and environmental settings, and instance-based transfer modeling experiments were conducted in five representative small-sample regions (A-E). Results show that the synergistic strategy significantly improved modeling performance in most regions, with maximum increases in predictive power (as indicated by R 2 ) of up to 18% compared with the baseline global transfer model. Remarkably, even when the number of global samples was reduced from 20,961 to around 200, the proposed strategy still outperformed local modeling and conventional global modeling approaches. In relatively stable environments, higher weights on environmental similarity yielded the best models, whereas in highly heterogeneous regions, spectral similarity played a more dominant role. The synergistic strategy also optimized the distribution of important spectral bands, enhanced SOC-responsive features in the visible region (450–750 nm), suppressed redundant information, and improved modeling efficiency. This study demonstrates that the proposed spectral-environmental synergistic sample transfer modeling method not only challenges the conventional assumption that “more samples guarantee better models” but also provides a novel pathway and theoretical support for the efficient use of global soil spectral libraries in regional SOC modeling.