LAUSR

Abstract Near infrared (NIR) spectroscopy is a potential technique for the quantification of the temperature reached (TR) in burned soils. Due to spatial variation, inaccurate predictions can result from calibrating a model with heat-sensitive compounds that are not present in the samples of the burned area. Therefore, we investigated how to develop robust models. The progressive augmentation of the model size successively enhanced the precision, while the increase of the calibration set's variability gradually improved the accuracy through decreases in bias. The increase in calibration set variability enhances the probability of calibration using only the most common heat-sensitive compounds, facilitating reliable predictions of TR regardless of the spatial variation. On the other hand, models calibrated with heated aliquots from a unique sample, even from a composite sample, should be totally avoided because, regardless of their apparent utility, they are prone to inaccurate predictions.