The present invention relates to a non-direct measurement temperature-compensating model correction method in the on-line application of a near-infrared spectrum analyzer, which comprises: acquiring a near-infrared spectrum of each sample under different temperature levels; respectively carrying out preprocessing and principal component analysis on the acquired spectra for temperatures and to-be-measured physical property parameters; then merging the obtained spectra to generate new spectral data; using partial least squares to model the spectral data to obtain measured values at the current moment; finally, constructing an on-line recursive algorithm, and thereby on-line near-infrared measurement with a non-direct measurement temperature compensation function is fulfilled. The present invention adopts temperature as a separate latent factor variable to participate in the process of near-infrared modeling, so that when near-infrared measurement is used, by depending on the adaptability of a model to the temperature, physical property measurement under different temperatures can be carried out without direct temperature measurement information or related calculation, and thereby the created model has better universality. The on-line recursive algorithm of the present invention has better adaptability to sample temperatures and other measurement condition changes.