Since the unavailability of local reference freshness values at individual pixels, a direct validation is impossible for the spatial distribution of pork freshness, i.e., the freshness maps visualized through applying the chemometric models that are trained on average spectra of regions of interest (ROI) to the spectra at individual pixels within the ROIs. Therefore, a dual-criteria evaluation of the freshness maps that were produced through different chemometric systems coupled with varied spectral filtering on both accuracy and precision was proposed. The former was quantified by the coefficient of determination of prediction (R2P) and the root mean square error of prediction between the chemical reference of ROIs and the average of the predictions at all individual pixels therein. The latter was quantified with the ratio of the pixels having negative TVB-N values to those of the ROI for a given subject, since the non-negativity according to the theoretical range of the freshness measurement. A bank of drastically different freshness maps of the same batch of pork were produced by using partial least squares regression (PLSR), over the visual/near infrared spectral range over 550~970nm, both before and after spectral filtering using ideal average smoothing filters with five different bandwidths of 6nm, 18nm,30nm, 42nm, and 54nm, respectively. The full range of consecutive wavebands, as well as 20 or 6 feature bands which were selected by successive projection algorithm (SPA), were used to form a collection of 18 combinations of bandwidth and the number of spectral bands to build chemometric models. Drastic difference resulted between the 18 approaches to visualization of freshness distribution. Analysis result showed, however, that all freshness maps were of good accuracy, equal to that of the chemometric models despite the lower quality of the spectra at individual pixels, even after spectral filtering, than those used in the training of models. And the precision of spatial predictions of freshness seemed to be co-determined by both spectra quality at individual pixels and the waveband-gains of chemometric models, and dominated by the former, R=0.72. It may be concluded that the spatial distributive predictions from imaging chemometrics can be objectively evaluated according to the statistics of the local predictions at pixels and the theoretic range of quality-indicating attributes;accuracy of quality-indicating maps, predicted on spectra at pixels, would not change from that of a linear chemometric system;better precision of spatial distribution prediction could be expected if spectral signal-to-noise ratio at pixels was improved and a chemometric model’s gains of wavebands were low.