Effective canopy spectral information is confined to mixed pixels containing cotton canopy, soil, and shadow components in unmanned aerial vehicle (UAV) remote sensing imagery, leading to low inversion accuracy of cotton flavonol content (Flav). There are also various performances among different machine learning models and feature combinations. In this study, an inversion framework was developed for cotton Flav using UAV remote sensing. A series of optimizations were integrated with canopy-mask constraints, mixed-pixel decomposition, multi-type feature fusion, and machine learning. UAV visible-light and multispectral images were acquired at three flight altitudes of 50, 80, and 110 m, while cotton leaf Flav was synchronously measured in the field. Four mixed-pixel decomposition methods were compared for multispectral imagery, including the linear spectral mixing model (LSMM), the generalized bilinear model (GBM) incorporating interactions among endmembers, the L2-regularized generalized bilinear model (L2-GBM), and the LSMM-GBM hybrid model (Hybrid), under canopy-mask constraints, endmember sets, and high-purity pixel selection. A total of 10 visible-light vegetation indices, 8 multispectral vegetation indices, 9 color features, and 8 texture features were extracted subsequently. Pearson correlation analysis with collinearity diagnostics was used for feature screening. 6 vegetation indices, 2 color features, and 2 texture features were retained as sensitive variables for cotton Flav estimation. Seven feature combinations were constructed, including vegetation indices, color features, texture features, vegetation indices with color features, vegetation indices with texture features, color features with texture features, and the combination of vegetation indices, color features, and texture features. Six regression algorithms were employed to establish cotton Flav estimation models, including partial least squares regression (PLSR), random forest (RF), support vector regression (SVR), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost). The parameters of six algorithms were tuned using internal cross-validation in the training dataset during optimization, according to the structural characteristics of each algorithm. Furthermore, a comparison was made on PLSR parameter tuning with three optimization strategies, namely bagging ensemble learning, correlation-weighted adaptive optimization, and differential evolution (DE). The coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) of the test dataset were evaluated to determine the final inversion model. After that, the data were collected at the boll-opening stage in 2025, the flowering-boll stage in 2026, and the boll-opening stage in 2026. Cross-field, cross-growth-stage, and cross-year generalization tests were conducted to evaluate the generalization of the model under varying field, phenological, and annual conditions. The results showed that the mean cotton endmember abundance ranged from 0.9961 to 0.9973 at flight altitudes of 50, 80, and 110 m, with standard deviations from 0.0048 to 0.0061. The cotton endmember abundance after estimation was more stable than that of the GBM, L2-GBM, and Hybrid. There were strong correlations between the 18 vegetation indices extracted from the multispectral imagery and Flav after LSMM processing. Among them, 16 vegetation indices showed that the absolute Pearson correlation coefficients increased with Flav at all three flight altitudes. The average increase in the absolute correlation coefficients was 0.0635, 0.0714, and 0.0715 at 50, 80, and 110 m, respectively, indicating that mixed-pixel processing improved the response of vegetation indices to variations in cotton Flav. Among the six machine learning models, PLSR achieved the highest overall accuracy for Flav estimation. After parameter optimization, the cross-validation-tuned PLSR model (CV-PLSR) achieved an R² of 0.9433, an RMSE of 0.0339, and an MAE of 0.0266. The generalization tests with data from different growth stages and years showed that the R² of CV-PLSR ranged from 0.8401 to 0.9100, while RMSE ranged from 0.0289 to 0.0363 and MAE ranged from 0.0228 to 0.0296. Stable estimation performance was achieved in the field, growth stages, and years. The findings can provide technical support for rapid and non-destructive field monitoring of cotton flavonol content and physiological differentiation of cotton breeding materials.

Locating weed growth points is often required in laser weeding. However, the high accuracy is also confined to the small target size, irregular morphology, and mutual occlusion in cornfield environments. In this study, the apical meristem localization (SDPL) framework was proposed to integrate an improved deep learning detector with the skeleton and density peak analysis. Firstly, a lightweight model, named YOLOv11n-LBD, was developed to detect the small, irregular, and partially occluded weeds. Three enhancements were incorporated using YOLOv11n: a Lightweight Adaptive Extraction (LAE) module for efficient and adaptive feature encoding, a bidirectional feature pyramid network (BiFPN) for the multi-scale feature fusion with learnable weights, and a DyHeadDCNv4 detection head with dynamic attention and deformable convolution for refined spatial and scale-aware feature representation. The subtle features were then extracted under complex fields. A custom dataset was constructed to collect corn seedlings and five common weed species. A series of tests was also conducted to verify the framework. The results showed that the improved model achieved a mean average precision (mAP) of 94.8% and a recall of 93.2%. Compared with the baseline YOLOv11n, the parameter count and computational complexity were reduced by 0.4 million and 0.8 G, respectively, with a real-time inference speed of 79.4 frames per second. There was a favorable balance between accuracy and efficiency suitable for mobile or embedded deployment in field robots. The SDPL localization was introduced to reduce the positional deviation between the bounding box geometric center and the actual biological growth point of a weed. The region of interest (ROI) was cropped using the YOLOv11n-LBD output. A high-quality binary mask was then generated for noise suppression using Gaussian filtering, followed by Otsu's adaptive thresholding for robust segmentation, effectively isolating the plant from the background. Morphological opening and closing operations were applied to suppress noise for the high connectivity. Subsequently, a skeleton optimization was implemented. The initial single-pixel skeleton was extracted using the Zhang-Suen parallel thinning algorithm, thus preserving the topological structure. The raw skeleton was smoothed and refined to reduce the artifacts, such as jaggedness and spurious branches. A continuous and natural skeletal representation was obtained after B-spline curve fitting. Candidate growth points were defined as the junction points of the refined skeleton branches. Finally, the precise growth points were identified to evaluate the local pixel density distribution within a circular neighborhood around each candidate. The final localization points were selected as the highest local density peak, indicating the most concentrated area of plant tissue near the stem base. Experimental validation on field images showed that there was a growth point localization accuracy of 81.2%. The accuracies were improved by 3.0 and 14.0 percentage points, respectively, compared with the key point detection and bounding box center positioning. The SDPL performed reliably for weeds with irregular shapes under partial occlusion, and even when overlapping with crop plants, where the geometrical method often failed. Robust detection and morphological-densitometric analysis can provide a precise, practical, and computationally efficient visual perception in laser weeding robots, contributing to targeted and energy-efficient weed control in sustainable agriculture.

Sunflower is one of the preferred oil crops in the mildly to moderately saline-alkali soils, due to its strong tolerance to nutrient-poor soils and drought. Among them, sunflower disk rot can be caused by sclerotinia sclerotiorum infection, leading to increasing sterile seed rates. Nutritional components can also decline, such as kernel protein and oil content. Thereby, both the edible and commercial value of the seeds are often required for high yield. Consequently, it is crucial to efficiently and accurately detect disk rot severity for early disease control, precision pesticide application, and yield estimation. However, conventional disease classification has been limited to manual efficiency and subjectivity in recent years. In this study, an improved YOLOv12n-RCL model was proposed to detect the severity of sunflower disk rot at the mature stage. 1) C3K2-RC modules were used to replace C3K2 ones in the backbone and neck networks. Receptive-field attention convolution (RFAConv) and coordinate attention (CA) mechanisms were integrated to enhance the feature extraction from the different severity grades of disk rot in complex environments. 2) A lightweight upsampling operator, CARFAE, was integrated into the neck network for feature reconstruction. 3) A lightweight shared convolutional detection head with separated batch normalization (LSCSBD) was introduced to improve the detection accuracy and speed for small-scale lesions. Experimental results show that the YOLOv12n-RCL model achieved a precision, recall, mAP_0.5, and mAP_0.5~0.95 of 83.4%, 81.2%, 84.8%, and 50.2%, respectively, which was improved by 3.8, 3.2, 3.4, and 4.0 percentage points over the baseline model. The number of parameters, computational complexity, and model size were reduced to 2.11 M, 5.7 GFLOPs, and 4.5 MB, respectively, corresponding to reductions of 17.9%, 12.3%, and 19.6%, compared with the original model. The normalized confusion matrix indicated that the recall for the disease severity grades 0, 1, 2, 3, and 4 were 83.0%, 81.2%, 79.1%, 80.7%, and 82.0%, respectively, indicating a balanced recognition over all five grades. Furthermore, 8.2% of samples with a true label of grade 2 were misclassified as grade 3, and 9.3% of samples with a true label of grade 3 were misclassified as grade 2. Therefore, the confusion between severity grades occurred primarily between grades 2 and 3. The misclassification rate remained below 10% in the rest, indicating its strong performance for the disk rot at different grades. A field test was conducted on 70 images with 1 298 samples of different grades of disk rot. The YOLOv12n-RCL model achieved only 10 missed detections and 8 false detections. Compared with the baseline YOLOv12n model, the superior performance was achieved with a higher frame rate of 27.5 frames per second (FPS). Visualization was also conducted on the sunflower disk rot areas using UAV orthophotography. The improved model maintained stable performance even in complex scenarios with dense target distribution and leaf occlusion, without significantly missing or false detections. In summary, the YOLOv12n-RCL model improved the detection accuracy to effectively balance the lightweight deployment efficiency. The finding can also serve as an algorithmic reference to detect the sunflower disk rot at the mature stage in smart agriculture.

Film covering is one of the most important parts during cotton planting. Multiple advantages remain, such as the heat preservation and moisture retention, inhibition of weed growth, fertilizer utilization rate, as well as the reduction of pests and diseases for the high yield. Among them, the film of cotton seedlings has been limited to low film-breaking efficiency. However, the labor-intensive manual breaking cannot fully meet the needs of large-scale cotton planting. Intelligence and precision operations are often required for the cotton film breaking using artificial intelligence in recent years. Intelligent machines for the cotton film breaking can be expected to accurately and rapidly identify the target of cotton seedlings under the film. In this study, a lightweight detection model, YOLOv11n-PRML, was proposed to optimize using YOLOv11n as follows: (1) A PConv-CGLU hybrid module was proposed to combine the PConv (Partial Convolution) of FasterNet and CGLU (Convolutional Gated Linear Unit) of TransNext. The C3k2 module was refactored to reduce the model complexity for the feature extraction; (2) An RSCD (Rep Shared Convolutional Detection) head with a heavy parameter-sharing strategy was employed to improve the accuracy and processing speed of the model in small target detection tasks; (3) The loss function was optimized as MPDIoU (Minimum Points Distance Intersection over Union) to improve the detection performance in dense environments; (4) The model was lightweighted using the LAMP (Layer-Adaptive Magnitude-Based Pruning) strategy. The TIDE (Toolkit for Identifying Detection and Segmentation Errors) evaluation indicator was introduced to evaluate the performance. The superiority of the YOLOv11n-PRML model was verified to detect the cotton seedling under film. Ablation tests were carried out to compare the different models. The experimental results show that the YOLOv11n - PRML model attains a precision of 90.1% and a mean Average Precision 0.5 (mAP_0.5) of 89.6%, representing increases of 1.8 and 1.0 percentage points, respectively, compared to the original YOLOv11n model. Its detection speed is improved to 114.4 frames per second. The localization error, missed ground truth error, and model size are 0.83, 0.92, and 4.0 MB, respectively, showing decreases of 0.32, 0.85, and 1.5 MB compared to the original model. Compared to the models like YOLOv5s-S (YOLOv5s-ShuffleNetV2), YOLOv7-tiny-M (YOLOv7-tiny-MobileNetV3), YOLOv8n-G (YOLOv8n-GhostNetV2), YOLOv9t, YOLOv10n, and YOLOv12n, the model's mAP_0.5 increased by 0.5, 4.3, 0.1, 4.0, 4.8, and 1.6 percentage points, respectively, while the model size was reduced by 2.8, 6.8, 1.2, 0.3, 1.8, and 1.3 MB, respectively. The YOLOv11n and YOLOv11n-PRML models were deployed on the NVIDIA GeForce RTX 2070 Ti platform, respectively, and optimized by TensorRT high-performance operators and Int8 quantization technology. The test results show that the detection accuracy mAP_0.5 of the YOLOv11n model is 87.6%, and the detection speed is 60.7 frames per second, while the detection accuracy mAP_0.5 of the YOLOv11n-PRML model is 89.1%, and the detection speed is 80.3 frames/s. In terms of inference time, the inference time of the model before and after the improvement is 16.3 ms and 11.1 ms respectively. YOLOv11n-PRML outperforms YOLOv11n in all indicators on the mobile terminal. In conclusion, the proposed YOLOv11n-PRML model provides technical support for the detection of cotton seedlings under film in complex environments and facilitates its practical application in real-world scenarios. It also offers a valuable reference for the development of intelligent film-breaking machinery. Future research will focus on expanding the dataset to include seedlings from multiple cotton varieties and conducting extensive comparative experiments to enhance the model's versatility and robustness.