Accurate prediction of leaf nitrogen concentration (LNC) is crucial for optimizing nitrogen (N) management; however, the robustness of hyperspectral LNC prediction models often varies with management practices and growth conditions. To improve model robustness, this study developed a feature-level transfer learning framework that integrates the Maximum Mean Discrepancy (MMD) with Continuous Wavelet Transform (CWT). Five maize hyperspectral datasets from arid and semi-arid regions of Inner Mongolia, including multiple varieties, growth stages, and irrigation and nitrogen regimes, were used to identify representative LNC-sensitive spectral features. The proposed approach was evaluated against a conventional partial least squares regression with variable importance in projection (PLSR-VIP) baseline and implemented within a transfer learning framework to assess its robustness across datasets. Our results showed that, compared with the PLSR-VIP approach, the MMD-CWT framework was able to identify LNC-sensitive and stable spectral features across datasets with contrasting cultivars, irrigation regimes, and management practices, which were primarily located in the red and red-edge spectral regions. In spite of better performance of PLSR-VIP within single-site datasets, its predictive accuracy was worse when transferred across datasets with different crop varieties and management conditions. In contrast, the MMD-CWT framework provided more robust, and transferable LNC predictions (R2 = 0.57-0.81), even though there was a limited set of five wavelet-derived spectral features. These findings highlight the practical value of integrating MMD-based domain alignment and CWT-based feature selection within existing transfer learning frameworks for robust and transferable LNC monitoring across diverse management conditions.