Soil organic carbon (C) and nitrogen (N) contents have an essential role in soil fertility, but they may be affected by salinity, which is especially responsible for land degradation in arid and semiarid regions. The objective of this work was to study the ability of visible and near infrared diffuse reflectance spectroscopy (VNIRS) to predict soil C and N contents and electrical conductivity (EC, a proxy for soil salinity) in variably salt-affected topsoils of the Sine Saloum region (Senegal). Different calibration procedures and spectral pretreatments were compared, and variable log-transformation usefulness was evaluated for prediction optimization.Predictions involved three calibration procedures: global partial least squares regression (PLSR), which used all calibration samples similarly; locally weighted (local) PLSR, with target samples predicted individually by giving higher weight to closest calibration spectra; and global PLSR per salinity class, after spectral discrimination of these classes. Predictions were performed with possible spectrum pretreatments (e.g., derivatization) and variable decimal log-transformation.The study was performed on 311topsoil samples (0–25 cm depth), either unsalted to slightly salty (Salt-, EC ≤ 2mScm−1; 262samples) or medium to highly salty (Salt+, EC > 2mScm−1; 49samples). Soil salinity was accurately discriminated using spectra: in validation, 100% and 95% of Salt- and Salt+ samples were correctly assigned on average, respectively. Best C and N content predictions were achieved after log-transformation using calibration by class (R2VAL = 0.87) and local calibration (R2VAL = 0.77), respectively; best EC prediction was achieved without log-transformation using global calibration (R2VAL = 0.90). This suggested C and N content predictions were affected by salinity; logC and logN distributions were almost symmetrical, hence log-transformation usefulness, while logEC distribution was very asymmetrical. No pretreatment yielded systematically good predictions; nevertheless, first-order derivative using 31-point gap often yielded good predictions, and second-order derivatives poor results.
Read full abstract