Automated large-scale farmland preparation operations face significant challenges related to path planning efficiency and uniformity in resource allocation. To improve agricultural production efficiency and reduce operational costs, an enhanced method for planning land preparation paths is proposed. In the initial stage, unmanned aerial vehicles (UAVs) are employed to collect data from the field, which is then used to construct accurate farm models. For single-field operations, a path planning approach is developed that minimizes energy consumption. The approach combines the selection of optimal operational angles with the implementation of efficient turning strategies, aiming to achieve full coverage. In addressing the issue of scheduling multiple machines across multiple fields, a two-stage optimization method, referred to as the BNSGA-III algorithm, is introduced. This algorithm integrates the NSGA-III algorithm with Beluga Whale Optimization (BWO), adaptive parameter adjustment, and Adaptive Inversion Crossover (AIC). The proposed method tackles the inherent complexity of agricultural environments, balancing operational efficiency and resource allocation through multi-objective optimization. Experimental results demonstrate that, compared to random operation directions, the proposed method reduces the path length by 1.9% to 3.1%, decreases the turning frequency by 19.5% to 24.0%, and improves coverage by 1.0% to 1.4%. In the context of multi-machine scheduling, the BNSGA-III algorithm outperforms the NSGA-II, NSGA-III, and MOEA/D algorithms, achieving improvements in total travel distance (12.3% to 34.4%), path balance (60.9% to 66.2%), and workload distribution (78.7% to 92.9%). Further evaluation shows that BNSGA-III excels in key metrics such as convergence (IGD), solution quality (HV), and diversity (Spread), thereby confirming its superiority in solution quality, convergence, and diversity. The findings of this study provide strong support for the advancement of intelligent agriculture.

Climate change inevitably affects vegetation growth in the Tibetan Plateau (TP). Understanding the dynamics of vegetation phenology and the responses of vegetation phenology to climate change are crucial for evaluating the impacts of climate change on terrestrial ecosystems. Despite many relevant studies conducted in the past, there still remain research gaps concerning the dominant factors that induce changes in the start date of the vegetation growing season (SOS). In this study, the spatial and temporal variations of the SOS were investigated by using a long-term series of the Normalized Difference Vegetation Index (NDVI) spanning from 2001 to 2020, and the response of the SOS to climate change and the predominant climatic factors (air temperature, LST or precipitation) affecting the SOS were explored. The main findings were as follows: the annual mean SOS concentrated on 100 DOY–170 DOY (day of a year), with a delay from east to west. Although the SOS across the entire region exhibited an advancing trend at a rate of 0.261 days/year, there were notable differences in the advancement trends of SOS among different vegetation types. In contrast to the current advancing SOS, the trend of future SOS changes shows a delayed trend. For the impacts of climate change on the SOS, winter T_max (maximum temperature) played the dominant role in the temporal shifting of spring phenology across the TP, and its effect on SOS was negative, meaning that an increase in winter T_max led to an earlier SOS. Considering the different conditions required for the growth of various types of vegetation, the leading factor was different for the four vegetation types. This study contributes to the understanding of the mechanism of SOS variation in the TP.

Rotational irrigation grouping changes network flow paths and distal pressures; empirical grouping or nominal-flow balancing therefore may not ensure hydraulic feasibility. For an automated cotton drip-irrigation system in Xinjiang, China, we coupled a field-calibrated EPANET model with CP-SAT in a closed loop. Type-specific minimum operating pressure heads for short- and long-hose units were derived from field pressure–discharge responses. CP-SAT generated candidate groups subject to fixed group size, outlet-hydrant mutual exclusion, branch-load balancing, and high-risk-unit dispersion. EPANET checked each group, and every complete infeasible nine-unit combination was returned as a no-good constraint before re-optimization. After calibration, model MAE, RMSE, and \(R^2\) were 0.264 m, 0.346 m, and 0.842; matched-condition holdout validation yielded 0.391 m, 0.488 m, and 0.727, respectively. The adopted pressure heads were 4.6 m for short-hose units and 4.9 m for long-hose units. All 19 final groups met the adopted type-specific pressure criteria in steady-state hydraulic simulations, while the field-tested groups achieved a pressure-satisfaction rate of 97.22%. The dripper-discharge coefficient of variation, Christiansen uniformity coefficient, and low-quarter distribution uniformity in the field-tested groups were 8.06%, 93.02%, and 89.58%, respectively. Under the observed operating conditions, feedback removed low-pressure combinations missed by nominal load balancing.

The cotton field weed detection model must balance detection accuracy, model size, and inference efficiency when deployed at the edge. This study constructed a dataset comprising four types of weeds based on field images collected from cotton fields in Xinjiang and proposed a lightweight detection model, RGC-YOLO, based on YOLOv8n. The model incorporates R-UIB modules into the deep layers of the backbone network to reduce the number of parameters and computational load during the high-channel feature extraction stage. A G-C2f module is introduced at the P5 feature fusion point in the neck network to improve the efficiency of deep semantic feature fusion. Additionally, a DIoU loss function is employed to strengthen the constraints on bounding box localization. The results show that, compared to YOLOv8n, RGC-YOLO achieves a 2.65% improvement in the mAP@0.75, while reducing the number of parameters and GFLOPs by 35.22% and 15.37%, respectively. This indicates that the model effectively reduces complexity while enhancing high-IoU localization performance. Eigen-CAM analysis further demonstrates that RGC-YOLO produces more concentrated feature responses in weed target regions and reduces irrelevant background activations. On the Jetson AGX Orin platform, the FP16 TensorRT engine for RGC-YOLO achieves high throughput and a compact footprint, demonstrating that its lightweight design translates into benefits for edge deployment. External validation on a sesame-field binary crop–weed dataset and the 2SeasonWeedDet8 multi-class weed dataset further confirms that RGC-YOLO maintains comparable detection performance while preserving its lightweight advantage across different agricultural scenarios.

To address the challenges of crop–weed instance segmentation and image-plane operation-point localization under visual similarity, background interference, leaf occlusion, and irregular plant morphology in cotton fields, a lightweight instance-segmentation model, GDMM-YOLO11, was developed. Based on YOLO11n-seg, the four backbone stage-transition downsampling convolutions at P2/4, P3/8, P4/16, and P5/32 were replaced with GhostConv to reduce redundant computation, a C3k2_DySnakeConv_Mona module was introduced to strengthen structural and multi-scale feature representation, and the original post-SPPF C2PSA block was replaced with mixed local channel attention (MLCA) to recalibrate high-level feature responses. Experiments were conducted on 2177 images containing 2856 annotated plant instances using a stratified 65%/15%/20% training–validation–test split. Across three independent runs with random seeds 3407, 3408, and 3409, GDMM-YOLO11 achieved a mask precision of 89.82 ± 2.60%, mask recall of 84.50 ± 2.49%, mask mAP@0.5 of 89.76 ± 0.66%, and mask mAP@0.5:0.95 of 68.50 ± 0.34%. Relative to YOLO11n-seg, the corresponding three-run mean values were numerically higher by 3.52, 1.07, 1.67, and 2.33 percentage points, respectively, while the parameter count decreased from 2.836 M to 2.533 M and the computational cost decreased from 9.6 to 8.8 GFLOPs. For image-plane operation-point localization, 505 of 528 ground-truth weed instances obtained valid same-class mask matches. The proposed skeleton-constrained fused-center method achieved a ground-truth-mask inclusion rate of 99.41%, a conditional Success@0.15 of 91.29%, and an end-to-end Success@0.15 of 87.31%. Paired comparisons with the mask-centroid baseline showed statistically significant improvements in ground-truth-mask inclusion and weed-boundary clearance, whereas differences in localization error and Success@0.10/0.15 were not statistically significant. TensorRT FP16 deployment on an NVIDIA Jetson AGX Orin achieved a model-only inference latency of 2.796 ± 0.115 ms, corresponding to 357.67 FPS. These results show that GDMM-YOLO11 provides a favorable accuracy–complexity trade-off while supporting image-plane operation-point generation and high-throughput model-only edge inference.

Precision weeding in cotton fields requires joint plant detection and growth-point localization despite canopy overlap and limited embedded computing resources. We propose ADQ-PACR, a one-stage framework that combines quality-aware localization with auxiliary coordinate regularization. ADQ jointly predicts growth-point coordinates, keypoint validity, and scale-normalized localization quality. PACR maintains independent supervision of the primary candidate and supervises the auxiliary candidate with the lowest coordinate error. The auxiliary branches are removed before deployment, leaving a fixed primary path. Experiments used 3527 smartphone images collected in Shaya and Changji, Xinjiang, from April to July 2022, spanning cotton seedling to squaring stages and including weeds such as Amaranthus retroflexus, Chenopodium album, and Portulaca oleracea. Evaluation used 343 fixed test images and three training seeds, with a common evaluator for adapted YOLO pose baselines initialized from detection weights and two-stage systems using predicted boxes. The full ADQ-PACR framework increased mean keypoint mAP50–95 over the corresponding YOLOv8n, YOLO11n, and YOLO26n baselines by 3.29, 2.56, and 2.17 percentage points, respectively. The YOLO11n implementation achieved box mAP50–95 of 60.68 ± 0.11% and keypoint mAP50–95 of 43.50 ± 1.55%. TensorRT FP16 deployment on Jetson AGX Orin in MAXN mode yielded end-to-end latencies of 20.85–22.22 ms across the three backbones.