Xiaoxiong Sun ^* Liangkuan Zhu Dayang Liu

• Northeast Forestry University, Harbin, China

Mechanical damage reduces the value of fruits. Therefore, early detection of fruit damage is crucial. The focus of this paper was on proposing a non-destructive detection method for early mechanical damage in blueberries (variety: Sapphire), based on hyperspectral image fusion combined with a multi-strategy improved support vector machine (SVM) model. Firstly, spectral features and image features of blueberry hyperspectral information were extracted using successive projections algorithm (SPA) and Grey Level Co-occurrence Matrix (GLCM), respectively. Secondly, SVM, RF and PLS-DA models based on spectral, image, and fused information were established, and these models were compared and analyzed. Finally, the hyperparameters of the SVM model based on feature fusion were optimized using a multi-strategy improved Beluga Whale Optimization (BWO) algorithm, and the classification accuracy of the SVM models with unoptimized, BWO-optimized, and multi-strategy improved BWO-optimized hyperparameters was compared and analyzed in order to determine the optimal model for early detection of blueberry damage. The results indicated that the SVM model established using feature fusion information achieved the highest classification accuracy upon being optimized by the multi-strategy improved BWO algorithm. The classification accuracy in the test set was 95.00%. Overall, the fusion of hyperspectral image information demonstrated high efficiency in the field detection of early bruising in blueberries. However, it required stringent conditions regarding the detection environment, such as light intensity and temperature. This model demonstrated the potential for the application of detecting early bruising in blueberries post-harvest.