Objective Ground-based hyperspectral remote sensing has shown strong potential for classifying wheat cultivars' nitrogen use efficiency (NUE). Construction of parameter system of the remote sensing technology for precise identification of wheat NUE can provide technical support for nitrogen management of the wheat production.

Methods Field experiments were conducted during the growing seasons of wheat in 2022–2023 and 2023–2024 at the Yuanyang Experimental Station of Henan Agricultural University. A total of 45 wheat cultivars that are widely grown in the Huang-Huai wheat region were chosen as the experiment materials. Each cultivar was subjected to normal nitrogen supply (N+) and no nitrogen supply conditions (N−), with three replicates. At the jointing, booting, flowering, grain filling, and maturity stages, hyperspectral reflectance data of the canopy were acquired using a ground-based Analytical Spectral Devices spectrometer, and plant samples were collected to determine the biomass and nitrogen content of stems, leaves, and spikes. At maturity, grain yield and grain number per spike were investigated. The nitrogen physiological use efficiency, nitrogen translocation efficiency, and total plant nitrogen accumulation at maturity were then calculated, resulting a total of 8 indicators for the subsequent analysis. Principal component analysis (PCA) was used to construct a comprehensive NUE index, and K-means clustering was applied to classify wheat cultivars into nitrogen-efficient, moderately nitrogen-efficient, and nitrogen-inefficient groups. Continuous wavelet transform (CWT) and first derivative transformation (DIFF1) were applied to preprocess the hyperspectral data. Subsequently, the ReliefF algorithm and the CARS method were used for feature selection. Wheat NUE classification models were developed using k-nearest neighbors (KNN), random forest (RF), and partial least squares discriminant analysis (PLS). Model performance was evaluated using overall accuracy (OA) and Cohen’s Kappa coefficient.

Results The coefficients of variation for the eight nitrogen-related indices ranged from 14.31 to 231.25%. The first two principal components explained 85.799% of the total variance in 8 N-related traits. Based on K-means clustering, 29 cultivars were classified as nitrogen-efficient, 28 as moderately nitrogen-efficient, and 33 as nitrogen-inefficient, establishing a rapid and reliable agronomic evaluation method for wheat classification. When classification models were constructed using full-spectrum reflectance, classification accuracy increased with advancing growth stages, reaching the highest performance at the grain-filling stage (mean OA=67.6%, Kappa=0.505), while the lowest accuracy occurred at the jointing stage (mean OA=58.8%, Kappa=0.327). Compared with DIFF1 and original reflectance (OR), CWT-based models improved classification accuracy by 2.9% and 6.2%, respectively. Under different feature selection strategies, CWT-based models outperformed DIFF1- and OR-based models by 3.3% and 7.3%, respectively. The ReliefF feature selection method effectively reduced the influence of weakly correlated bands, and models constructed using ReliefF-selected features (mean OA=67.6%, Kappa=0.508) outperformed those based on CARS-selected features (mean OA=66.1%, Kappa=0.452) and full-spectrum models (mean OA=62.8%, Kappa=0.419). Among classification algorithms, KNN achieved the highest classification accuracy (mean OA=79.7%, Kappa=0.662), whereas PLS exhibited the lowest performance (mean OA=65.8%, Kappa=0.489).

Conclusion Based on the eight agronomic traits at grain filling stage, the constructed wheat nitrogen-efficient cultivar classification model using the CWT + ReliefF + KNN approach demonstrates high accuracy (OA=85.2%, Kappa=0.745), showing it as a reliable technical framework for the rapid identification of nitrogen-efficient wheat cultivars using ground-based hyperspectral remote sensing.