High accuracy and low computational complexity can be required to detect the tea pest and disease in modern agriculture. In this study, a lightweight and high-precision model, called WGSE-YOLOv11n, was proposed using an improved YOLOv11n architecture. A tea pest and disease dataset was also constructed to support the model training and validation. The dataset comprised 9 categories (healthy leaves, tea anthracnose, tea leaf spot disease, tea black rot, tea leaf rust, tea leaf blight, tea white spot disease, tea aphids, and tea spider mites), totaling 2 496 sample images. Two parts of the data sources consisted of: One part comprised 746 valid images that were captured at the Yunfeng Ecological Tea Garden in Hanzhong, Shaanxi Province, China, in 2025. Another part consisted of 1 750 images that were sourced from the Roboflow public dataset. All images were uniformly resized to a resolution of 640×640. Label Studio was used to annotate the disease regions in the images and then convert them into YOLO format. Simultaneously, Mosaic data augmentation (including translation, rotation, scaling, brightness adjustment, and noise injection) operations were applied to expand the sample size to 6 150 images. The dataset was ultimately divided into the training (4 296 images), validation (1 235 images), and test (619 images) sets at a ratio of 7:2:1. In the WGSE-YOLOv11n model design, firstly, the Wavelet-Gaussian Dynamic Convolution (WGD) architecture was developed in the backbone network. The frequency-domain decomposition of the WaveletPool module was integrated with the lightweight properties of the C3k2-GhostDynamic module. Feature representation was enhanced to reduce the parameter complexity, thereby improving the capture of multi-scale lesion features. Secondly, the neck network was incorporated with a Wavelet Pooling Star Fusion (WSF) architecture. WaveletUnPool upsampling and StarFusion feature enhancement modules were combined to effectively reconstruct the lesion structural and texture information, thereby improving the recognition accuracy for the minute lesions and complex boundaries. Finally, a lightweight EfficientHead detection module was introduced. Channel grouping via the GroupConv architecture further reduced the parameters and computational complexity. Grad-CAM was employed to generate the heatmaps for the tea leaf pest and disease detection, in order to validate the performance of the detection. Results show that the WGSE-YOLOv11n model accurately and rapidly located the lesions, whose heatmaps exhibited the high-response zones in the pathological tissue boundaries, indicating the strong spatial coupling with the lesion morphology. In contrast, the heatmaps of the YOLOv11n baseline model shared significant spatial diffusion, indicating the weak responses in the diseased areas, edge attenuation, and false activation of non-pathological tissues. Edge response intensity was significantly enhanced in the WGSE-YOLOv11n model, while the false-positive activations were reduced substantially. There was no feature confusion in the multi-leaf scenarios. The better performance of the improved model also outperformed that of the baseline, particularly in the response intensity toward the minute targets, like the ring lesions of the tea cake disease and tea aphids. There were great improvements in the pathological feature focus, edge precision, and multi-class robustness. The improved model was also deployed on a Jetson Orin NX development board that connected to a D455 camera for real-time image capture. The practical performance on the mobile devices was further validated after deployment. Among them, TensorRT was integrated for operator acceleration and INT8 quantization. CUDA was then utilized for the multithreaded parallel preprocessing. The computational efficiency enhanced the detection rate, in order to prevent computational constraints on the embedded platforms. Experimental results demonstrate that the WGSE-YOLOv11n model achieved 97.64% precision, 97.87% recall, and 99.08% mean average precision on the self-built dataset, respectively, indicating the improvements of 0.24, 3.62, and 0.77 percentage points respectively, over the baseline YOLOv11n model. Parameter count, computational load, and model size were compressed to 1.51 million parameters, 3.3 gigabytes of computational load, and 3.2 megabytes, respectively, which were significantly reduced by 41.5%, 47.6%, and 39.6%, respectively, compared with the baseline model. On Jetson Orin NX, the detection frame rate reached 26.14 frames per second, with an average inference time of 14ms per image. The high accuracy of the recognition was obtained to substantially reduce the computational load, suitable for the deployment on mobile and embedded devices, as well as the real-time detection in field environments.

Automated harvesting holds the promising potential in the informatization and automation of smart agriculture. Manual harvesting can be replaced to improve the high efficiency and cost saving. However, the existing models are limited to the large number of overlapping citrus fruits, branch and leaf occlusion in the complex environments, particularly for the much more parameters and high computational complexity. In this study, an improved YOLOv8-MEIN model was proposed for the citrus recognition using YOLOv8n. Firstly, the more efficient (ME) convolution module was designed to improve the CSPDarknet53 into 2-Stage feature pyramid network (C2f) module of YOLOv8n. The high number of common convolutional parameters and computation were reduced to fully meet the requirements of the real-time detection, due to the conditions of mobile hardware performance. The overall computation was also reduced to decrease the number of parameters in the model. Secondly, the Inner- complete intersection over union (CIoU) loss function was used to accelerate the bounding box regression for the better performance of the model. The weak generalization and slow convergence of the CIoU loss function were also compensated in the detection task. The reason was that the CIoU loss function only considered the shape loss in original YOLOv8, thus degenerating to the IoU loss function, when the target and bounding box were at the same aspect ratio. Finally, the YOLOv8-MEIN model was validated for the citrus detection. The YOLOv5s, YOLOv7-tiny, Faster R-CNN, and YOLOv8n models were selected for the training and testing on the citrus dataset. A comparison was carried out on the self-built citrus dataset under the same number of iteration rounds, in order to verify the performance of the YOLOv8-MEIN model. The YOLOv8-MEIN was improved 5.85 percentage points in the recall, 14.3 percentage points in the mean average precision (mAP) mAP0.5 metrics, and 34.8 percentage points in mAP_0.5~0.95 metrics, compared with the two-stage Faster R-CNN, where the model size was only 5.1% of Faster R-CNN. Compared with the lightweight networks of YOLO series, such as YOLOv5s, YOLOv7-tiny, and YOLOv8n, the number of parameters was reduced by 59.1%, 52.3%, and 4.4%, respectively, whereas, the mAP_0.5~0.95 values were improved by 0.7, 4.0, and 0.6 percentage points, respectively, indicating the outstanding mAP_0.5. The experimental results show that the YOLOv8-MEIN model was achieved in the higher detection accuracy and efficiency with the smallest model volume and the lower computational cost. More convenient model was found for the transplantation and application suitable for the real-time target detection of citrus. The size of model and the number of parameters were reduced by 3.3 and 4.3 percentage points, respectively, compared with the mainstream target detections. The improved model was fully met the requirements for the fast and accurate recognition of citrus fruits in the complex environments. The finding can provide the technical references for the automated harvesting of citrus fruits.

Timely and accurate identification of leaf diseases can greatly contribute to the effective pest prevention and control in the tomato growth cycle. In this study, a lightweight detection approach was proposed to balance between detection accuracy and computational efficiency. An enhanced version of the YOLOv8n (you only look once) model was augmented with a pruning algorithm, specifically designed to identify the various types of tomato leaf diseases. The YOLOv8n architecture was used to replace the conventional C2f (concatenated feature fusion) module with the more efficient StarBlock module. The number of parameters were was significantly reduced within the network. The complex features were represented to improve the overall expressive power of the improved model. Additionally, a mixed local channel attention mechanism (MLCA) was integrated to capture the richer set of contextual information and multi-scale features, which were critical to distinguishing among different types of leaf diseases. Furthermore, multi-level channel compression was used to reengineer the original detection head. The performance of the improved model was refined to reduce the dimensionality of the feature maps along the channel axis. Thus, the a more streamlined and computationally efficient structure was obtained after these architectural adjustments. Sparse training was then performed on the high-precision and lightweight model. Some weights were selectively eliminated within the network, according to their importance. A specified level of sparsity was achieved after training. Experimental results indicate that there was the best trade-off between data sparsity and model performance at a sparsity rate of 0.005. The redundant or less significant channels were removed using channel pruning. The final high-precision and lightweight models were obtained after training. A series of tests were carried out to validate the improved model. A dataset was comprised of 4,130 images. Nine distinct types of tomato leaf diseases were compiled. The developed model was then tested against this dataset. Compared with the baseline YOLOv8n, the improved model exhibited a substantial reduction in the parameter count (63.3%), floating point operations (72.8%), and model weight size (61.9%). The better performance of the improved model was achieved in the accuracy (97.5%), recall (96.2%), and average precision (mAP@0.5: 98.5%), with the a minimal average drop in the performance metrics of just 0.23 percentage points. Furthermore, the improved model was deployed on mobile devices, indicating the a remarkable detection frame rate of 358.5 frames per second, with an average inference time of 4.4 milliseconds per image. Once benchmarked against the popular object detection frameworks, such as Faster R-CNN and SSD, the improved model demonstrated a dramatic decrease in both the number of parameters (by 207.3 and 34.9 M, respectively) and computational complexity (by 497.2 and 280.6 G, respectively). The YOLOv8n was reduced by 20.0% and 59.3% in the parameters, corresponding to a 17% and 49.54% decrease in the computational complexity, compared with the more recent and lightweight models, like YOLOX Nano and YOLOv5Nano. The competitive accuracy levels were all preserved. In conclusion, the improved YOLOv8-SLMP model can offer a highly viable solution to the real-time detection of tomato leaf diseases. Particularly, the footprint was also minimized, in terms of the parameters, computations, and model size. An ideal candidate was deployed on the resource-constrained embedded systems, thus enabling more widespread and efficient monitoring of crop health.