In this study, hyperspectral imaging technology with a wavelength range of 450 to 1000 nanometers was used to collect spectral data from 160 Korla pear samples at various maturity stages (immature, semimature, mature, and overripe). To ensure high-quality data, multiple preprocessing techniques such as multiplicative scatter correction (MSC), standard normal variate (SNV), and normalization were employed. Based on these preprocessed data, a custom convolutional neural network model (CNN-S) was constructed and trained to achieve precise classification and identification of the maturity stages of Korla pears. Additionally, a BP neural network model was used to determine the characteristic wavelengths for maturity assessment based on the sugar content feature wavelengths. The results demonstrated that the BP model, based on sugar content feature wavelengths, effectively discriminated the maturity stages of the pears. Specifically, the comprehensive recognition rates for the training, testing, and validation sets were 98.5%, 93.5%, and 90.5%, respectively. Furthermore, the combination of hyperspectral imaging technology and the custom CNN-S model significantly enhanced the detection performance of pear maturity. Compared to traditional CNN models, the CNN-S model improved the accuracy of the test set by nearly 10%. Moreover, the CNN-S model outperformed existing techniques based on partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM) in capturing hyperspectral data features, showing superior generalization capability and detection efficiency. The superior performance of this method in practical applications further supports its potential in smart agriculture technology, providing a more efficient and accurate solution for agricultural product quality detection. Additionally, it plays a crucial role in the development of smart agricultural technology.

The effective classification methods of the small target objects in the no-fly zone are of great significance to ensure safety in the no-fly zone. But, due to the differences of the color and texture for the small target objects in the sky, this may be unobvious, such as the birds, unmanned aerial vehicles (UAVs), and kites. In this paper, we introduced the higher layer visualizing feature extraction method based on the hybrid deep network model to obtain the higher layer feature through combining the Sparse Autoencoder (SAE) model, the Convolutional Neural Network (CNN) model, and the regression classifier model to classify the different types of the target object images. In addition, because the sample numbers of the small sample target objects in the sky may be not sufficient, we cannot obtain much more local features directly to realize the classification of the target objects based on the higher layer visualizing feature extraction; we introduced the transfer learning in the SAE model to gain the cross-domain higher layer local visualizing features and sent the cross-domain higher layer local visualizing features and the images of the target-domain small sample object images into the CNN model, to acquire the global visualizing features of the target objects. Experimental results have shown that the higher layer visualizing feature extraction and the transfer learning deep networks are effective for the classification of small sample target objects in the sky.

Fruit occlusion, dense clustering, platform vibration, and cross-frame re-identification errors reduce the reliability of video-stream yield estimation in high-density dwarf apple orchards. To address these limitations, this study developed a vision-based framework for apple detection, tracking, dynamic counting, and yield estimation. Apple detection was improved by embedding a Shuffle Attention (SA) module into YOLOv12n and introducing SlideLoss for hard-sample reweighting, yielding the YOLOv12n-SA-SlideLoss detector. Following detection, the DeepSORT tracking module was also improved by augmenting it with global motion compensation (GMC) and a dual-criteria association strategy based on motion consistency and appearance similarity. Yield estimation was formulated as a counting-based compensation model using dynamic fruit counts, mean fruit mass, and an occlusion compensation factor, and was compared with a pixel-geometric regression model based on detection boxes. Field experiments were conducted in a high-density dwarf ‘Fuji’ apple orchard in Aksu, Xinjiang, China. YOLOv12n-SA-SlideLoss achieved an mAP@0.5 of 95.7%, exceeding the YOLOv12n baseline by 2.3%. The improved DeepSORT reached 87.4% multiple object tracking accuracy and 86.7% multiple object tracking precision, while reducing identification switches by 22.2% relative to the original algorithm. The dynamic counting framework increased average counting accuracy from 85.6% to 93.5%. For yield estimation, the counting-based compensation model achieved a mean prediction accuracy of 85.2%, an RMSE of 18.5 kg, and a mean bias of −17.2 kg across six validation samples (n = 6), each covering multiple trees. Overall, the results indicate that combining detection, motion-compensated tracking, and counting-based correction improves stability of single-view video-stream yield estimation in high-density dwarf apple orchards.