Partial least-squares regression models were developed using spectra of tomato fruit collected at 15°C and tested on spectra of an independent set of fruit at higher sample temperatures. The influence of sample temperature on the model used to predict fruit dry matter( DM) was manifested primarily in terms of bias, not standard error of prediction. For example, a model for DM created with samples at 15°C had a bias of –0.9% DM and –1.9% DM when used to predict DM in fruit at 25°C and 35°C, respectively. The addition of spectra of a relatively small number of samples collected at different temperatures to the calibration set can create a model that is robust to temperature; however, continued addition of samples at a uniform temperature overwhelms this compensation effect, resulting in a model that is not robust to temperature. For a model that included spectra of fruit at a range of temperatures, the prediction bias increased as the ratio of samples at 15°C to samples at other temperatures increased beyond 200:1. The use of orthogonal scatter correction, external parameter orthogonalisation, generalised least-square weighting, global model development and repeatability file were compared for the development of a temperature-robust DM model. The use of a repeatability file is recommended on the basis of the lowest root mean square error of prediction and bias. Selection of wavelength regions to avoid water absorption features is recommended for an attribute not associated with an OH feature (such as skin-colour prediction).