Criação e atualização de modelos de calibração multivariada para determinação de carbono orgânico no solo por espectroscopia NIR

Abstract

The organic carbon is an indicator of soil quality and currently its quantification is made by the modified Walckley & Black method, which in addition to spend a lot of time, use reagent and then generates toxic waste. Thus, it is possible to use an alternative method based on multivariate calibration model. The model was constructed using a linear method of partial least squares (PLS), 1000 samples of different Brazilian soils and 17 latent variables. Outliers were identified and removed. This model was used to quantify the organic carbon in an external samples set and the results were not satisfactory indicating that the current model needs to be updated to ensure its durability. For this, strategies were used to update the PLS model with representative samples of the external set, the DS and PDS standardization algorithms were apply to verify its ability in the PLS model uptdate. The updated models performance were analyzed by comparing the prediction errors (RMSEP) and correlation coefficient to adjust of external samples (Rv), it was observed that the inclusion of a percentage of representative samples of the external set is the best strategy to update the model.