Operationalizing soil spectral libraries: a case study for soil carbon in peat soils of Switzerland

Infrared spectroscopic modeling can link and massively scale up traditional laboratory methods for many soil characteristics in a cost-effective and timely manner. This has prompted the development of soil spectral libraries (SSLs) from the regional to global scale. The operational value of SSLs lies in their ability to provide estimations of key soil properties on a scalable basis, from field to regional levels. While various approaches have been proposed for utilizing SSLs for estimating local soil characteristics, such as spiking, subsetting, memory- or instance-based learning, or transfer learning, few studies have focused on organic soils. Given that peat soil conservation is a top priority in Europe, we developed and tested three spectroscopic modeling strategies in the mid-infrared range to estimate soil carbon at local target sites in Swiss peatlands with minimal sampling and measurement requirements. We compared the accuracy of local models, models using the Swiss SSL in conjunction with local samples, and subsets of local and Swiss SSL samples using a data-mining algorithm called resampling local (RS-LOCAL). Our findings indicate that local models require 50 samples for accurate results, with a root mean squared error (RMSE) < 3% total carbon. Models utilizing the entire Swiss SSL with local samples yielded biased predictions with lower accuracy than local models. However, subsets of local and Swiss SSL samples using RS-LOCAL only required 5 or 10 local samples to accurately predict soil carbon (RMSE = 3.16% or 2.71% total carbon, respectively). Thus, we recommend using RS-LOCAL together with a SSL as an efficient and accurate strategy to tailor soil characteristic estimations for a specific region. Even though the Swiss SSL was under-representative of organic soils, adequately mining the information in the SSL was sufficient for predicting soil carbon in independent peatland regions. This will aid in reducing field and laboratory work and scaling up the acquisition of quantitative soil information over space and time.