Using Digital Soil Mapping to identify where best to store soil carbon

Abstract

Soil has the potential to store large amounts of soil organic carbon (SOC). The mineral surface area (MSA - m2 g-1) indicates this capability. Given the cost of laboratory analysis and the heterogeneous nature of soil, a digital soil mapping (DSM) approach may enable a farmer to determine where best to sequester SOC. Herein we create DSM’s of MSA at the Lincoln University Dairy Farm (LUDF) on the South Island of Aotearoa (New Zealand). We compare ordinary kriging (OK) of topsoil (0-0.15 m) MSA, with DSM’s created using various digital data either or alone or in combination, including remote (i.e., Light Detection and Ranging [LiDAR]) and/or proximal (i.e., gamma-ray [γ-ray] and electromagnetic [EM] induction) sensed data. We compare various approaches; stepwise multiple-linear regression (MLR), geostatistical (i.e., regression kriging [RK]), and a hybrid machine learning and geostatistical (i.e., Random Forest [RK] with RK [RFRK]) models. In all cases, a prediction sample size (n = 160-10) was evaluated against a validation (n = 40) dataset, considering prediction agreement (Lin's concordance correlation coefficient [LCCC]) and accuracy (root mean square error [RMSE]). We also compare which digital data is best and whether used on its own or in combination. When n = 160, OK produced moderate (Lin’s = 0.78) agreement. When predicting with digital data (i.e., g-ray [K], EM [1mPcon] and DEM [SWI]), RK (0.88), RFRK (0.82) and MLR (0.80) produced substantial agreement. When considering the minimum number of samples, OK maintains near substantial (i.e., > 0.8) agreement when n = 100 (0.79), while RK (0.84) and RFRK (0.8) need n = 20, with MLR (0.79) as few as n = 10. We conclude that with digital data (i.e., SWI, g-ray K and 1mPcon) and using RK was optimal, with as few as 1 sample required every 8 hectares for calibration, but we recommend 1 sample be taken every 4 hectares. Moreover, to improve prediction additional soil samples should be taken where MSA was large (i.e., > 30 m2 g-1). When considering which digital data might be useful on its own or in various combinations, we found that in terms of agreement and considering models developed using RK, using DEM + EM (0.88) was best in terms of agreement, followed by EM only and DEM, g-ray + EM (0.86), with DEM only (0.83) and DEM + g-ray (0.81) also having substantial agreement. The final DSM of RK, indicate where soil capability would be greatest to store SOC, and this was where the most suitable soil conditions (i.e., MSA) were present in the southern part of the LUDF, associated with the larger clay content of the Glay soil profiles.