Abstract: Based on near-infrared spectroscopy, an in-situ detection method for the corrosion of the steel core in power transmission lines was proposed. Firstly, the near-infrared spectra of the surfaces of four types of power transmission lines with different corrosion states were obtained through near-infrared detection. Then, the best spectral data preprocessing method was obtained through comparative analysis. The optimal wavelengths for modeling were selected by the latent projection graph (LPG). A recognition model of corrosion state classification based on support vector regression (SVR) was established by combining principal component analysis (PCA) dimension-reduced data and Pelican optimization algorithm (POA) optimized parameters. Finally, the accuracy of the corrosion state recognition model was verified using energy spectrum analysis data. The results show that a maximum variance explained rate of 99.16% could be achieved using standard normal variable processing and Savitzky-Golay smoothing preprocessing. And 10 optimal wavelengths were selected out by LPG. The visual clustering results of the 4 classes of samples were obtained by combining the best spectral data preprocessing method with the best wavelengths and using PCA. The final classification accuracy rate was 96.43% by inputting PCA 2D score data into POA-SVR classification model.