Abstract: 【Objective】This study aimed to establish a the method for determining nitrogen in protected-cultivation lettuce based on Terahertz spectroscopy and imaging， identify the optimal prediction model for nitrogen content in protected-cultivation lettuce， so as to provide theoretical and technical support of applying terahertz spectroscopy technology in agricultural product quality analysis and plant nutrition detection， thereby achieving nitrogen regulation and fertilizer optimization in precision agriculture.【Method】Using protected-cultivation hydroponic lettuce as the research objects， lettuce plants were cultivated for 40 d under four nitrogen concentration gradients of 20%， 60%， 100%， and 150%. Actual values of nitrogen content were determined using the Kjeldahl method. A terahertz spectral imaging system was employed to acquire time-domain spectral images of lettuce leaves at different nutrient levels. The Savitzky Golay （S-G） smoothing algorithm and multivariate scatter correction （MSC） were used for spectral information preprocessing. Feature bands were then identified using stability competitive adaptive reweighted sampling （sCARS）， interval partial least squares （iPLS）， and iterative information retention variable （IRIV）. Based on this， a prediction model of lettuce nitrogen content was established by combining Liner-Kernel， RBF-Kernel， and KELM kernel functions. The best prediction model was selected through coefficient of determination （R²） and root mean square error （RMSE）.【Result】0.2-1.2 THz turned out to be the optimal frequency range for analyzing nitrogen content in lettuce. Through dimensionality reduction analysis and feature variable selection using sCARS， iPLS， and IRIV algorithms， 6， 17， and 15 feature variables were selected in the power spectrum dimension， while 7， 16， and 12 feature variables were selected in the absorbance dimension， respectively. In the two dimensions of power spectrum and absorbance， three different kernel functions were integrated with feature variable extraction algorithms for modeling， and the results showed that the overall modeling effect in the power spectrum dimension was better than that in the absorbance dimension. Among them， the LS-SVM prediction model exhibited the best prediction performance in extracting feature variables using the sCARS algorithm of RBF kernel power spectral dimension. The R² and RMSE of the calibration set were 0.9640 and 0.1939， respectively， and the R² and RMSE of the prediction set were 0.9606 and 0.1986， respectively. The KELM prediction model also achieved the best prediction performance by extracting feature variables using the sCARS algorithm in the power spectrum dimension. The R² and RMSE of the calibration set were 0.9675 and 0.1913， respectively， while the R² and RMSE of the prediction set were 0.9596 and 0.1997， respectively.【Conclusion】Terahertz spectroscopy shows recognizability for changes in nutrient levels in plant tissues， which may be closely related to tissue structure， polarization behavior， and macromolecular vibration patterns caused by nitrogen changes. Therefore， a nondestructive nitrogen detection method based on terahertz spectral imaging can effectively predict the nitrogen content in lettuce， and the LS-SVM prediction model has the best prediction effect by extracting feature variables using the RBF nuclear power spectral dimension sCARS algorithm.