Prediction and error modelling of soil organic matter based on spectral reflectance and geostatistical stochastic simulation.

Soil organic matter (SOM) is a key property in evaluating soil degradation. Diffuse reflectance spectroscopy is an alternative approach to conventional methods for soil analysis. This paper was aimed at developing a prediction model of SOM based on laboratory measurements of spectral reflectance within the 350-2500 nm spectral range and presenting a modelling of SOM error based on geostatistical stochastic simulation. The study area was the Turbolo watershed (Calabria, southern Italy) representative of areas highly prone to soil degradation. Two hundred four topsoil samples were collected and each sample was used for both spectroscopic measurements and laboratory determination of SOM content. The partial least squared regression (PLSR) analysis was used on only 152 samples to establish the relationships between spectral reflectance and SOM. The optimum number of factors to retain in the calibration models was determined by cross validation. The models were independently validated using the other 52 soil samples. Results revealed a high level of agreement between measured and predicted values. Five hundred simulations of SOM error were generated using conditional sequential Gaussian simulation algorithm and statistical information was extracted: 1) the map of the 'expected' error and that of standard deviation; 2) the probability maps of overestimation and underestimation.