Early detection of plant diseases with automated, non destructive and high-throughput techniques is a major objective in plant breeding and crop protection. Near infrared spectroscopy and hyperspectral imaging are proven to be particularly relevant technologies. However, robust discriminant models remains a challenge because of the many uncontrolled sources of variability during the experiment. Indeed, at early stages of most diseases, the temporal variations due to environment and measurement effects can induce signal shifts of greater magnitude than the infection itself, masking the information of interest. Excluding the variations of the measurement environment and the temporal fluctuation of the plant-pathogen interaction can depreciate the model robustness. Here, the problem is addressed in a study of the seven potato cultivars monitored for the presence of early blight disease at 0, 18, 36, and 96 h after inoculation. Three practical corrections are proposed regarding the effect of temporal fluctuations. (i) subclass effect, (ii) kinetic effect of the disease, and (iii) measurement effect. Eventually, the application of EPO-PLSDA to orthogonalise the model regarding temporal variation to produce invariant models proved to be the only suitable and well-performing of the tested solutions. With this approach the disease can be detected from 36 h after inoculation for 6 of the 7 tested cultivars. Classification errors differ among the cultivars but on average are below 25% of errors.