Non-destructive early bruise detection of blueberries based on hyperspectral image feature fusion PKOA-SVM model

To achieve non-destructive detection of early post-harvest bruising of blueberries, this study proposes a detection method using hyperspectral image feature fusion PKOA-SVM model. Firstly, by comparing various preprocessing methods, the first-order and second-order derivative transformation combined method (FD-SD) was determined as the optimal spectral preprocessing method. On this basis, the high-spectrum spectral features and multi-dimensional spectral map information such as image color, texture, and shape were fused to construct a multi-source feature set. To eliminate the influence of dimensionality, the Z-score method was used to standardize the feature set. Then, the competitive adaptive reweighting algorithm (CARS), the continuous projection algorithm (SPA), and the genetic algorithm (GA) were used to optimize the standardized spectral features and combined with the texture features extracted by the gray-level co-occurrence matrix (GLCM), a high discriminative fused feature set was finally formed. Finally, the peacock optimization algorithm (PKOA) was introduced to automatically optimize the key hyperparameters (penalty factor C and kernel function parameter γ) of the support vector machine (SVM), and a PKOA-SVM classification model was established. Experimental results show: (1) FD-SD preprocessing can effectively improve the model performance; (2) Among various feature combinations, the "texture + spectrum" fusion strategy has the best effect, with a detection accuracy of 97.91% for the 24-hour fruit stalk group (an increase of 5.61%, versus 92.30% for individual spectral features) and 91.67% for the 30-minute flower petal group (an increase of 8.34%, versus 83.33% for individual spectral features); (3) Compared with other optimization algorithms, PKOA exhibits better optimization efficiency and stability, and the optimized SVM model in the 24-hour fruit stalk group and 30-minute flower petal group reaches 98.92% and 94.56% respectively. This research method provides an effective solution for high-precision non-destructive detection of early bruising of blueberries, and can provide method references for intelligent detection of agricultural product quality. • Mechanical bruising during postharvest handling accelerates blueberry decay and compromises fruit quality. • A novel method combining hyperspectral-image feature fusion with PKOA-optimized SVM is proposed for early bruise detection. • Multimodal feature fusion outperforms single spectral features, with texture-spectral fusion achieving optimal performance. • The PKOA-optimized SVM model demonstrates enhanced accuracy and robustness in early bruise classification.