The chlorophyll fluorescence Fv/Fm parameter has great significance in plant stress. Current acquisition approaches for the plant’s Fv/Fm need dark adaptation, which cannot realize a real-time measurement. In order to achieve realtime acquisition of Fv/Fm, peppers under four water stress levels were used as research objects, and the Fv/Fm parameter was predicted based on hyperspectral imaging and characteristic wavelength screening methods. The median filter was used to denoise the Fv/Fm image, and the hyperspectral image was matched with the Fv/Fm image based on two-dimensional coordinate transformation. The three types of spectral preprocessing algorithms, including the standard normal variate (SNV), multivariate scattering correction (MSC), and Savitzky-Golay convolution smoothing (SG) were compared, and the successive projections algorithm (SPA) was used to select the characteristic wavelengths. Based on the optimal SG preprocessing algorithm, the modelling accuracy of the partial least square regression (PLSR), analytical error backpropagation (BP) neural network, and radial basis function (RBF) neural network were compared, and the BP algorithm showed the relatively high determination coefficient of 0.918 and root mean square error of 0.011 in the test set. In summary, the SG-SPA-BP modelling method reduced the complexity of the model, while maintaining a high prediction accuracy, which provided an effective approach for predicting the chlorophyll fluorescence Fv/Fm image in real-time.