In-field hyperspectral imagery is a promising tool for crop phenotyping or monitoring. In association with partial least square regression (PLS-R), it allows building high spatial resolution maps of the chemical content of plant leaves. However, several optical phenomena must be taken into account, due to their influence on collected spectral data. The most challenging is multiple scattering, produced when a leaf is partly illuminated by light reflection or transmission from neighboring leaves. It can induce bias in prediction results. This paper presents a method for multi-scattering correction. Its development has been based on simulation tools: a 3D canopy model of winter wheat was combined with light propagation modeling, in order to simulate the apparent reflectance of every visible leaf in the canopy for a given actual reflectance. Leaf nitrogen content (LNC) prediction has been considered. A data set of reflectance spectra associated with LNC values has been issued from real leaf measurements. A theoretical disturbance subspace representing the spectrum dispersion in the spectral space due to multi-scattering has then been built by considering polynomial combinations of the initial spectra, and a projection along this subspace has been applied to every simulated spectra. Using this strategy, a PLS-R model built on initial spectra was still satisfactory when applied to simulated spectra with multiple scattering. The method has then been applied to real plants in greenhouse and field conditions, and its prediction results compared with those of a standard PLS-R, confirming its efficiency in the presence of various lighting environments.