Peatlands ecosystem is one of the largest global terrestrial carbon pools. However, there is a shortness of its characterisation and information through new proximal sensing approaches. The visible and near-infrared spectroscopy is an inexpensive, quick, non-evasive, proximal sensing and low-cost analysis employed in field and/or laboratory. Despite that, there is another current issue in using this tool for creating global models, which is how it can retrieve local characteristics such as soil organic carbon (SOC) and total nitrogen (TN) in peatlands ecosystems. The aims in this study were to: (i) create a local model for quantifying SOC and TN finding the best pre-processing and machine learning methods in peatlands ecosystem, and (ii) evaluate the contribution of SOC and TN data collected in that ecosystem to global models in European Union. The hypothesis was that the SOC and TN data sampled in peatlands ecosystem can improve analytical quantification of those soil properties. The soil and spectral datasets were retrieved from the Land Use/Cover Area frame Statistical Survey with 21,771 observations at 0-20 cm depth and 63 soil cores in a degraded peatland in Germany with 262 observations up to 2 m depth. We evaluated three spectral pre-processing techniques with the Partial Least Square Regression (PLSR), Random Forest (RF), and Cubist machine learning algorithms. The best pre-processing technique was achieved applying Savitzky-Golay smoothing with a window size of 71 points, 2nd order polynomial, and zero derivative with Cubist algorithm for both SOC and TN predictions. Furthermore, merging the local with global data for global modelling demonstrated to improve SOC and TN predictions because of the local data representativeness and quality. Therefore, the SOC and TN data sampled in peatlands ecosystem can improve quantification of those soil properties in field and laboratory, which are crucial proxies for GHG emissions and climate change.
Keywords: Climate change; Greenhouse gases emissions; Machine learning; Proximal sensing; Soil spectroscopy.
Copyright © 2022 Elsevier Ltd. All rights reserved.