Non-destructive testing can be expected to rapidly and accurately detect the zinc content in oilseed rape leaves. In this study, a high-precision detection was realized to combine with deep transfer learning using hyperspectral imaging technology. Oilseed rape with similar growth shape was divided into two groups (Group Z and Group ZS) by soilless cultivation, each of which included the five types of stress reagents. Therefore, 400 oilseed rape leaf samples were selected for each type of stress reagent, while 2 000 oilseed rape leaf samples were collected for each group, leading to 4000 oilseed rape leaf samples in total. The hyperspectral image information of oilseed rape leaf samples was obtained by hyperspectral imaging equipment. The whole blade was taken as the region of interest. The average spectral information was obtained in the region of interest after calculation. Firstly, the predictive performance of different pre-treated spectra was compared for the zinc content in oilseed rape leaves under the action of silicon. Standard normalized variable (SNV) was established as the best pre-processing. The spectral data was processed by SNV for further analysis. A stacked auto-encoder (SAE) was used to reduce the dimensionality of the best pre-processed spectral data, compared with the traditional. Finally, transfer stacked auto-encoder (T-SAE) was performed on the optimal SAE deep learning network. The transfer learning model was obtained to verify the portability between the deep learning models in silicon-free and silicon environments. The results showed that the support vector machine regression (SVR) model with SAE extraction depth features shared the best prediction on zinc content in the oilseed rape leaves under silicon-free or silicon environments. The best performance was achieved in the SNV-SAE-SVR model under a silicon-free environment. The coefficient of determination (R_c²) and root mean square error (RMSEC) of the calibration set were 0.939 3 and 0.018 22 mg/kg, respectively, while the coefficient of determination (R_p²), root mean square error (RMSEP), and residual predictive deviation (RPD) of prediction set were 0.850 7, 0.034 66 mg/kg and 2.607, respectively. The R_c², RMSEC, R_p², RMSEP, and RPD of the prediction set were 0.963 4, 0.012 97 mg/kg, 0.876 6, 0.028 54 mg/kg and 2.732, respectively. In addition, the SVR model with T-SAE extraction depth features performed the best prediction on zinc content in both silicon-free and silicon environments, where the R_c², RMSEC, R_p², RMSEP, and RPD of the optimal SNV-T-SAE-SVR model prediction set were 0.970 5, 0.012 04 mg/kg, 0.881 0, 0.027 48 mg/kg and 2.966, respectively. Deep transfer learning combined with hyperspectral imaging technology can effectively detect the zinc content in oilseed rape leaves under both silicon-free and silicon environments.