In order to improve the accuracy of integrated soil-crop total nitrogen detection in agricultural fields, the canopy spectra of winter wheat were used as a research object to quantify the accuracy of four data reduction methods (neighborhood preserving embedding (NPE), t-distribution stochastic neighbor embedding (t-SNE), Laplacian eigenmaps (LE) and locally linear embedding (LLE)) in canopy spectral feature extraction and crop and soil total nitrogen content detection. The canopy spectral reflectance and the corresponding crop and soil total N contents of four varieties of winter wheat, namely Yumai 49-198, Zhoumai 27, Aikang 58 and Xinong 509, were collected at four levels of N application, respectively. The NPE, t-SNE, LE and LLE were used to downscale the data in the visible and partial near-infrared bands from 400nm to 900nm, and subsequently, a random forest regression model was developed based on the four sets of downscaled features. Comparison of the full-spectrum information and the prediction performance of the four sets of downscaled features for crop and soil total nitrogen content showed that the hybrid LLE-RF method achieved the best nitrogen prediction results with an R 2v value of 0.9150 for the coefficient of determination of crop total nitrogen content prediction and a root mean square error (RMSEP) of 0.2212mg/kg for crop total nitrogen prediction. The coefficient of determination R 2v for prediction of total soil nitrogen content was 0.8009. The RMSEP was only 0.0085mg/kg, which were all better than that of the original full-spectrum data as well as the other three sets of downscaled features. The experimental results showed that the LLE downscaled spectral information can effectively characterize the crop total nitrogen content and soil total nitrogen content.