Improved Soil Organic Carbon Prediction in a Forest Area by Near-Infrared Spectroscopy: Spiking of a Soil Spectral Library

Abstract

The rapid quantitative assessment of soil organic carbon (SOC) is essential for understanding SOC dynamics and developing management strategies in forest ecosystems. Compared with traditional laboratory methods, visible and near-infrared spectroscopy is an efficient and inexpensive technique widely used to predict SOC content. Herein, we compared three different spiking strategies. That is, a large-scale global soil spectral library (global-SSL; 3122 samples) was used as the basis for predicting SOC content in a small-scale local soil spectral library (local-SSL; 89 samples) in Wugong Mountain, Jiangxi Province, China. Partial least squares regression models using global-SSL ‘spiking’ with local samples did not necessarily achieve more accurate predictions than models using local-SSL. Using the developed strategy, a calibration set can be established by selecting the top N spectral samples from global-SSL with high similarity to each local sample, together with the ‘spiking’ set from local-SSL. It is possible to individually improve the prediction results based on local samples (R2 = 0.90, RMSE = 7.19, RPD = 3.38) and still allow for quantitative prediction from fewer local calibration samples (R2 = 0.83, RMSE = 8.71, RPD = 2.68). The developed method is cost-effective and accurate for local-scale SOC assessment in target forest areas using a large soil spectral library.