Citrus yield estimation is often required for low missed detection rates, tracking, and counting errors using UAV remote sensing, particularly for dense fruit occlusion and small target size. In this study, citrus tracking and counting were proposed using UAV remote sensing video imagery with improved YOLO11. The video data was first collected by a DJI Phantom 4 UAV at an angle of approximately 45°. A citrus target detection was constructed on the tracking dataset. An automatic citrus counting was realized using UAV video streams. A lightweight YOLO11-PMSL model was combined with an improved ByteTrack algorithm. The detection head layer was simplified to a three-level structure (P2-P4) using the YOLO11n network architecture. The feature pyramid structure was reconstructed. Deep redundant modules were removed to fuse the high-resolution shallow features, and then perceive small targets. Secondly, the C3k2-MSEIE multi-scale edge module was introduced after adaptive scale fusion and contour enhancement. The local details were expressed to extract the fruit contours. The overall morphological features of the fruit were preserved, with better feature expression in the densely populated fruit areas. Subsequently, the loss function was replaced by SIoU. A direction-sensitive constraint was introduced to improve the localization accuracy and training stability of the detection boxes. Finally, the LAMP was used to prune the model for the removal of redundant weights. The number of parameters and floating-point operations was then reduced to compress the model size for model lightweighting. The ByteTrack algorithm framework was improved, rather than using IoU in spatial location measurement. The accuracy and stability of fruit tracking were further enhanced in complex orchard environments. The similarity metric in ByteTrack was replaced with the DIoU. Simultaneously, a region-counting anti-shake mechanism was embedded in the algorithm. The target ID jump problem under occlusion was effectively solved for the accurate counting of citrus fruits. Experimental results showed that the YOLO11-PMSL model effectively improved the performance of the model. Specifically, compared with the original YOLO11n object model, better performance was achieved in the feature pyramid into a P2-P4 three-level structure. The number of model parameters was reduced to 116 m, the model size was compressed to 2.6 MB, and the recall and mAP0.5 metrics were significantly improved by 7.9 and 5.1 percentage points, respectively. The more lightweight model was verified by the higher accuracy for small targets. The precision, recall, and mAP0.5 were improved by 2.2, 10.7, and 8.7 percentage points, respectively, with the C3k2-MSEIE edge module. Once the loss function was replaced from CIoU to SIoU, the convergence speed was accelerated to further improve its performance. The LAMP algorithm was used to prune the model, fully meeting the lightweight deployment of edge terminals. The performance remained at the baseline level before pruning. While the number of parameters, floating-point operations, and model size were significantly reduced from before. Ultimately, the precision, recall, and mAP0.5 were improved by 3.3, 11.6, and 9.3 percentage points, respectively, in the object detection task. In terms of lightweighting, the number of parameters, model size, and floating-point operations were reduced by 86.05%, 76.36%, and 26.98%, respectively, compared with the original model. The detection speed was improved by 65.18%. The YOLO11-PMSL model achieved a detection accuracy and speed on the citrus dataset. In the object tracking task, the ByteTrack multi-object tracking algorithm achieved an accuracy of 92.8% and a tracking precision of 81.7%. Compared with the SORT, DeepSORT, and BotSort algorithms, the tracking accuracy was improved by 5.5, 5.7, and 4.3 percentage points, respectively, and the tracking precision was improved by 19.2, 19.4, and 10.8 percentage points, respectively. The average accuracy of citrus counting reached 88.4%, compared with manual counting. The counting error was smaller than that of manual counting. Citrus counting was effectively realized in farmland scenarios. This finding can provide a technical approach for citrus yield prediction.

Blast disease has been one of the most serious threats to rice production. However, the conventional evaluation of rice blast resistance cannot fully meet the large-scale production in recent years. In this study, an efficient evaluation was proposed for rice neck blast resistance via combining unmanned aerial vehicle (UAV) low-altitude remote sensing with the YOLOv7 model. 2 565 images of rice neck blast were collected using UAV and then divided into small-sized images (≤1240×1240 pixels) in the annotated area. The small-sized images were further subjected to a random combination of five operations, including rotation, scaling, translation, cropping, and changing contrast. The images with low resolution were removed after data cleaning. Finally, the dataset size was expanded with better diversity. The squeeze extinction attention and deformable convolution were introduced into the YOLOv7 model, in order to capture the fine-grain features of the rice neck blast disease spot. YOLOv7_Neckblast model was established for the rice neck blast detection. The number of affected ears was obtained for 15 rice varieties. The incidence rate was calculated for the disease grade of rice neck blast. Among them, 4, 4, 3, 5, 7, and 9 rice varieties of grades 1, 3, 5, and 7, as well as 9, 2, and 2 were assessed by YOLOv7_Neckblast, respectively. At the intersection over the union (IoU) threshold of 0.5, the mean average precision (mAP) of YOLOv7_Neckblast for rice neck blast was 66.4%, which was 4.0, 6.4, and 5.8 percentage points higher than that of the original YOLOv7, FCOS (fully convolutional one-stage object detection), and RetinaNet models, respectively. The recall rate was 75%, which was 4.0, 8.0, and 22.0 percentage points higher than those of the three models, respectively. The F1 score was 71%, which was 4.0, 5.0, and 12.0 percentage points higher than those of the three models, respectively. The training YOLOv7_Neckblast was relatively stable with the low floating-point operations per second (FLOP) under the same training conditions. The loss values remained stable with about 0.019 and 0.020 at the end of training, which was lower than those of FCOS and RetinaNet models. Furthermore, the molecular-assisted marker selection (MAS) showed that the contribution of the Pit and Pib genes to the resistance to the neck blast was 100.0% and 57.14%, respectively. The rice varieties carrying the Pit and Pib genes also exhibited a stronger resistance to disease. In addition, YOLOv7_Neckblast achieved an average accuracy of 86.67% in evaluating 15 rice varieties' resistance, compared with the actual resistance level. The low-altitude UAV remote sensing coupled with machine learning can be used to evaluate the resistance to rice neck blast for rice breeding.

Salt stress is considered one of the primary threats to cotton production. Although cotton is found to have reasonable salt tolerance, it is sensitive to salt stress during the seedling stage. This research aimed to propose an effective method for rapidly detecting salt stress of cotton seedlings using multicolor fluorescence–multispectral reflectance imaging coupled with deep learning. A prototyping platform that can obtain multicolor fluorescence and multispectral reflectance images synchronously was developed to get different characteristics of each cotton seedling. The experiments revealed that salt stress harmed cotton seedlings with an increase in malondialdehyde and a decrease in chlorophyll content, superoxide dismutase, and catalase after 17 days of salt stress. The Relief algorithm and principal component analysis were introduced to reduce data dimension with the first 9 principal component images (PC1 to PC9) accounting for 95.2% of the original variations. An optimized EfficientNet-B2 (EfficientNet-OB2), purposely used for a fixed resource budget, was established to detect salt stress by optimizing a proportional number of convolution kernels assigned to the first convolution according to the corresponding contributions of PC1 to PC9 images. EfficientNet-OB2 achieved an accuracy of 84.80%, 91.18%, and 95.10% for 5, 10, and 17 days of salt stress, respectively, which outperformed EfficientNet-B2 and EfficientNet-OB4 with higher training speed and fewer parameters. The results demonstrate the potential of combining multicolor fluorescence–multispectral reflectance imaging with the deep learning model EfficientNet-OB2 for salt stress detection of cotton at the seedling stage, which can be further deployed in mobile platforms for high-throughput screening in the field.