To rapidly detect Perilla aldehyde (PAE) content in Perilla Frutescens, hyperspectral imaging technology was employed, and the hyperspectral images of Perilla Frutescens were acquired from four distinct producing regions. Based on obtaining effective wavelength images of Perilla Frutescens, its texture features and energy values obtained by wavelet transform were fused with spectral values in various ways to create different characterization vectors. These vectors were then employed to construct corresponding rapid detection models for PAE content. The models' prediction capabilities were thoroughly compared and analyzed to determine the optimal prediction strategy for PAE content. The specific methods are as follows. (1) Four methods were employed to preprocess the raw spectral values. After evaluating the predictive performance of the constructed models, it was determined that Local Weighted Scatterplot Smoothing (LOWESS) emerged as the optimal preprocessing method. (2) The Competitive Adaptive Reweighted Sampling (CARS) and Successive Projections Algorithm (SPA) were employed to extract the characteristic wavelengths of the preprocessed spectral information, and the corresponding spectral values were then computed to facilitate their integration with other mentioned features in the paper. (3) Principal Component Analysis (PCA) was utilized to get effective wavelength images from the hyperspectral images. The grayscale co-occurrence matrix (GLCM) was then applied to the effective wavelength images to extract four texture features: Energy (ASM), Contrast (CON), Correlation (COR), and Entropy (ENT); simultaneously, the Daubechies wavelet was employed to conduct three-level decomposition of the effective wavelength image, and the energy of the low-frequency component derived from the decomposition was also considered as a characterization feature of the effective wavelength image. (4) The extracted features of wavelength spectral values, wavelet energy values, and texture features were utilized to construct feature input vectors in different ways, and based on the mentioned vectors, four detection models were then constructed: Partial Least Squares Regression (PLSR), Backpropagation Neural Network (BPNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost); and then these models were evaluated and compared according to their prediction capabilities to identify the optimal input vectors and prediction models. The research results indicate that the prediction capabilities of the four prediction models using single-class feature input vectors are all inferior to that of the input vectors fused with multi-class features; the optimal input vector is the feature fusion input vector, which incorporates the texture feature values and wavelet energy values from effective wavelength images as well as the spectral values corresponding to the feature wavelengths selected by CARS after LOWESS preprocessing. Among these models, the XGBoost model demonstrates the strongest prediction capabilities. The R-c(2) and RMSEC of the training set are respectively 0.998 08 and 0.022 49, and the R-p(2) and RMSEP of the testing set are 0.989 44 and 0.036 40, respectively. This research finding introduces a novel approach for the rapid detection of PAE content in Perilla Frutescens, and it also serves as a valuable reference for developing detection strategies for other components.
