To address the typical issue of insufficient labeled samples in the process of real-time total nitrogen (TN) detection, a low-dimensional space-based redundancy minimization semisupervised learning (LDSRM) framework is proposed. Unlike other graph-based semisupervised algorithms, the proposed method performs semisupervised modeling in a low-dimensional space, significantly reducing computational costs and improving algorithm performance. First, sparse feature selection is integrated into the initial graph model construction to preliminarily select key features from the original space and construct a low-dimensional space. Second, by minimizing the redundancy of features through the Pearson correlation coefficient, the negative impact of redundant features is reduced, and the most representative features are further selected in the low-dimensional space, thereby improving regression accuracy. Finally, we apply the proposed method to practical TN detection. The results demonstrate that this method can establish accurate detection models that meet national testing standards with a relative error below 10% even with limited labeled samples available.