Timely county-level estimation of wheat harvesting area was essential for cross-regional machinery dispatching, operational resource coordination, and emergency response during the concentrated harvest season. However, conventional monitoring workflows based on post-processing of agricultural machinery trajectories usually produced statistical results by noon of the following day, which limited their value for same-day decision-making. To address this time-lag problem, a rapid prediction method for county-level wheat harvesting area was developed by integrating heatmaps derived from operation trajectories with meteorological information. The study used operation trajectory data from the national agricultural machinery trajectory big data platform, county-level daily harvested area generated by standardized trajectory post-processing, and meteorological data derived from the European Centre for Medium-Range Weather Forecasts ERA5-Land dataset. For each county and day, operation trajectory points were converted into heatmaps through kernel density estimation. A Residual Network-152 model was then used as a fixed feature extractor after removal of the final fully connected classification layer, and deep spatial features were obtained from the heatmaps. These heatmap features were fused with cumulative precipitation, average air temperature, and average humidity aggregated up to 11:00 a.m. A random forest regression model was subsequently established to estimate the final cumulative harvested area of the same day from morning information. The model was trained and evaluated using data from 40 counties distributed across 8 major winter-wheat-producing provinces in China, namely Shanxi, Shaanxi, Henan, Hebei, Jiangsu, Anhui, Hubei, and Shandong. The observation window was uniformly defined from May 25 to June 25 in 2021–2023. Samples from 2021 and 2022 were used for model training, while samples from 2023 were used for testing. The results showed that the proposed method achieved stable prediction performance across all 40 counties. The coefficient of determination exceeded 0.6 in every county, with an average of 0.71 and a maximum of 0.91. The normalized root mean square error ranged from 0.06 to 0.21, with an average of 0.14. Better performance was observed in large and contiguous plain areas, where operation hotspots in the heatmaps were more coherent and easier to characterize. For example, the model performed particularly well in Quanjiao County, Feixi County, Wenshang County, Wudi County, Hua County, and Qing County. In contrast, prediction accuracy was relatively lower, though still acceptable, in fragmented hilly and mountainous counties such as Hengshan District, Lantian County, Yingshan County, Tongshan County, and Jianshi County. This decrease in accuracy was associated with more dispersed field patterns, weaker aggregation structures in the heatmaps, and possible heterogeneity in terminal coverage. The ablation analysis further clarified the contribution of each input source. When only heatmap features were used, the average coefficient of determination reached 0.62 and the average normalized root mean square error was 0.17, indicating that the heatmaps already captured most of the effective information related to harvesting intensity and spatial concentration. When only meteorological inputs were used, the average coefficient of determination dropped sharply to 0.07 and the average normalized root mean square error increased to 0.29, suggesting that meteorological conditions alone were insufficient to describe daily harvesting dynamics. When heatmap features and meteorological factors were combined, the average coefficient of determination increased to 0.71 and the average normalized root mean square error decreased to 0.14, confirming that meteorological information effectively complemented the operation-state information represented by the heatmaps. In addition, the sensitivity analysis on the Gaussian smoothing scale showed that the best overall performance was obtained when the smoothing parameter was set to 7 pixels, which balanced local noise suppression and preservation of spatial differences. Overall, the proposed method shifted county-level wheat harvesting area estimation from next-day statistics to same-day midday prediction while maintaining good agreement with trajectory-derived reference values. The result output time was advanced by about 20 hours, which demonstrated clear practical value for same-day machinery scheduling and rapid operational response during the wheat harvest season. At the same time, the results indicated that future improvements should further consider terminal coverage heterogeneity and complex terrain conditions to enhance model robustness in fragmented production regions.

Agricultural scenarios, such as tractor roads, village and farm paths, can often exhibit sparse geometric structure, weak vertical constraints, and high scene repetition. Since existing scene coordinate regression (SCR) methods are suitable for structured urban environments, they often suffer from unstable pose estimation in agricultural scenes. A single LiDAR frame can also contain insufficient discriminative information for accurate localization. This study aims to improve the localization accuracy and deployment efficiency of autonomous agricultural machinery using deep learning and LiDAR localization. AgriLoc was proposed as a temporal-enhanced LiDAR localization framework for rapid adaptation to new agricultural scenes with the high efficiency of frozen-backbone SCR pipelines. The localization strategy was selected to freeze a pre-trained scene-agnostic backbone and then train only scene-related prediction modules, thereby reducing the amount of scene-specific optimization during deployment. On top of this framework, AgriLoc also introduced a temporal sample classification guidance (T-SCG) and a multi-head attention enhancement module to exploit temporal context from consecutive LiDAR frames. Specifically, a sequence of LiDAR point clouds was processed to produce both global feature sequences and current-frame spatial dense features. A pre-trained classification head mapped the global features into region-discriminative guidance features, and then temporally encoded them by a GRU-based T-SCG module to capture motion continuity and suppress ambiguity caused by sparse or repetitive structures. A multi-head attention mechanism was applied to the full history of guided states. The forgetting effect of recurrent modeling over long sequences was further alleviated to adaptively focus on keyframes with stronger scene discriminability. The temporal context was fused with current-frame spatial dense features and then fed into the regression head to predict scene coordinates, which are then used by RANSAC to estimate the final 6-DoF pose. Experiments were conducted on a self-built dataset. A 9 km route was selected to cover the full operational chain of agricultural machinery, including departure from the machine shed, traversal through village roads and public roads, entry into farmland, field operation, and return to the shed. Two sequences were collected on the same route at time periods, with 30.4 GB (15 611 frames) and 23.0 GB (12 017 frames), respectively. Experimental results showed that AgriLoc achieved a mean translation error of 2.17 m and a mean rotation error of 2.75°. Compared with the baseline LightLoc, AgriLoc reduced the average translation error by 27.2% and the average rotation error by 6.4% in challenging tractor-road scenes, while only 3.8 h of training was required for a new scene. An ablation study further showed that the T-SCG reduced the mean translation error from 2.98 to 2.50 m, and multi-head attention was further improved to 2.17 m, indicating the complementary contribution of temporal guidance and attention after keyframe selection. Furthermore, cross-scene robustness of AgriLoc was evaluated on the NCLT public and campus dataset. The high cross-scene adaptability of AgriLoc was achieved, with a mean translation error of 1.88 m and a mean rotation error of 2.82° on NCLT, while 1.84 m translation error and 2.73° rotation error on the campus dataset. Overall, this finding can also provide a practical, high-precision, and rapidly deployable LiDAR localization for autonomous navigation of agricultural machinery.

The positioning environment of farm roads is complex and variable, with defined scenes including open sky, forest, overpass, and tunnel. Relying solely on the Global Navigation Satellite System (GNSS) for positioning throughout the operation of unmanned agricultural machines is insufficient. This study addressed the need for rapid and accurate identification of farm road scene types to select appropriate positioning devices for unmanned agricultural machines. Real-time satellite signal data, obtained through onboard GNSS receivers and combined with broadcast ephemeris data, were used to extract positioning status and track satellite statistics and distribution features. A classifier based on a sliding window was developed, and an inference based on tracked satellite numbers was proposed to detect the farm road positioning scene and calculate the length of different road segments. This study was conducted using real-world working conditions at an agricultural machinery cooperative in Miyun, Beijing, China, where three sets of GNSS data were collected from the farm road using a vehicle-mounted all-frequency GNSS receiver. The results showed that the classification accuracy for open-sky, overpass, and tunnel scenes was 93.96%, with a recall rate of 98.31% and an F1 score of 95.82%. Compared with random forest and XGBoost, the F1 scores improved by 17.83 and 5.70, respectively. This method, based on GNSS multi-feature fusion, can provide a reference for selecting multi-source positioning devices and planning the routes of unmanned agricultural machines.

Leveraging extensive trajectory data to analyze the operation modes of agricultural machinery for gathering precise spatial information is an important fundamental task for subsequent agricultural machinery trajectory research. However, complex algorithm models hinder nonspecialized researchers from further processing agricultural machinery trajectory data. In the present application note, ChatGPT is taken as an example and a complete prompt guide for large language models (LLMs) is provided for autonomously identifying the operation mode of agricultural machinery. This guide provides low-cost workflows for processing agricultural machinery trajectory data when computer science or data science expertise is lacking. It even possesses the capability to utilize newly learned algorithms such as the random forest model, which has not been previously explored in the literature for operation mode identification, to accomplish the task. To the best of our knowledge, this is the first attempt to apply LLMs to identifying agricultural machinery operation modes based on trajectory data. The complete prompt guide is publicly available at https://github.com/kakushuu/prompt-guide/.

Due to varying crop maturity periods and uneven distribution of agricultural machinery, China has developed a unique service model known as cross-regional agricultural machinery operations. Currently, China’s comprehensive mechanization rate for grain crops is relatively high, creating a substantial market for cross-regional agricultural machinery operations. Research on the behavioral patterns of cross-regional agricultural machinery migration is both urgent and significant. Considering the actual rules of cross-regional migration during the wheat harvest and the characteristics of the trajectory data, this paper proposes a trajectory mining method using a recursive search-based DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm. One representative finding of this study is that by mining the trajectory data of wheat harvesters within 25 d of peak harvest period, 131 cross-regional trajectories were identified, consisting of 11 633 harvesters. Three main routes of wheat harvester cross-regional migration were identified, along with several smaller routes outside their range. The overall spatiotemporal pattern aligns with observed realities in China. This study can provide valuable references for operators to optimize cross-regional routes, for agricultural machinery manufacturers to develop location-based services, and for relevant government departments to formulate policies.

Agricultural machinery is one of the most significant sources of non-road mobile emissions in recent years. Yet accurate estimation of its carbon and pollutant emissions is often hindered by the fuel consumption data and localized emission factors. In this study, a prediction framework was developed for tractor fuel consumption, CO₂, and NO_x emissions from tractors. Massive BeiDou trajectory data were also integrated with portable emission measurement system (PEMS) measurements and the national operational emission inventory. 1) The samples were incorporated with complete fields from the BeiDou trajectory data and multidimensional features, such as the tractor speed, direction, engine speed, and torque. A feature set was constructed to predict the fuel consumption. A feature engineering process was conducted on the dataset of trajectory samples with complete fields from the 2023 summer harvest. The trajectories were segmented into kinematic fragments with a time step of 100s. A Residual Deep Neural Network (ResDNN) model with a 21-layer structure was employed to predict fuel consumption, with an R² of 0.75 and a MAPE of 24.46%. Fuel consumption features were further constructed for the missing feature fields, according to the operational duration and distance. Various algorithms of machine learning were used for the predictive models. Light Gradient Boosting Machine (LightGBM) model performed best (R²=0.77, MAPE=39.28%). 2) Localized emission models were established. PEMS tests were then conducted on four typical China IV standard tractors with the rated powers of 51.5, 58.8, 88.2, and 177 kW under varying engine loads. According to the experimental data, machine learning algorithms (ridge regression, random forest, decision tree, SVM, and LightGBM) were compared to construct emission prediction models using multi-dimensional input features. The evaluation results indicated that the ResDNN model achieved the best performance for fuel consumption prediction with the features (R²=0.75, and MAPE=24.46%), effectively resolving the gradient degradation in deep networks. In the simplified feature set, the LightGBM model performed optimally (R²=0.77, and MAPE=39.28%). Support Vector Machine was identified as the optimal model for CO₂ emissions (R²=0.95, MAPE=5.98%) prediction, while LightGBM performed best for NO_x emissions (R²=0.86, MAPE=13.78%). Importance analysis revealed that instantaneous fuel consumption and engine speed were the most critical determinants for CO₂ emission, whereas environmental parameters shared a higher contribution to NO_x prediction. Finally, these models were applied to the trajectory data of 870,192 tractors equipped with BeiDou terminals during the 2023 summer harvest (May 25 to June 25). A high-resolution national emission inventory was estimated after application. The total fuel consumption was 47.01×106 L, thus generating 128.21×103t of CO₂ and 281.9t of NO_x. The average daily emissions per tractor were 147.33 kg for CO₂ and 0.32 kg for NO_x. Spatially, the emissions were concentrated in major agricultural provinces, such as Jiangsu, Henan, and Heilongjiang. Temporally, the emission peaks coincided perfectly with the busy farming rhythm. This finding can provide support data for carbon and pollutant emissions from large-scale agricultural machinery in green agriculture.

The target velocity jump has caused the low-velocity stability and fuel economy in the whole process of autonomous agricultural machinery operation. In this study, optimal velocity planning was proposed for the field operation of agricultural machinery in the constantly variable transmission (CVT) tractor. The global and local planning objectives and constraints were collected from the autonomous agricultural machinery. The minimum-jerk polynomial velocity planning was then designed using the Bellman optimality principle and optimal control theory. The velocity-following control system was also developed to realize the stable velocity cruise of autonomous agricultural machinery in the field. The velocity planning was modelled as multi-stage decision-making, according to the time series. The optimal decision was achieved in the jerk optimal control at each stage to realize the state transition. Bellman optimality principle and optimal control theory were selected to design the minimum-jerk polynomial velocity planning. The minimum jerk cost function was subject to hard constraints, such as the maximum and minimum velocity, acceleration, and jerk. The soft constraints were utilized to realize the driving task in the shortest possible time. Numerical solutions were used to reduce the difficulty of the model. Polynomial piecework fitting was generated to fully meet the smooth and continuous optimal speed reference curve under the hard constraints. The performance of the control system was significantly improved to prevent the velocity jump and varying conditions. The velocity stability and fuel economy were realized in the CVT tractor. The feasible solution space was obtained in the numerical solution. The end states of tractor motion were taken to heuristically search for the optimal solution under hard and soft constraints. The numerical solution was obtained after sampling. The unexecuted portion of the previous cycle’s solution in each planning cycle was utilized to achieve the time consistency of the unmanned tractor. The executive layer controller of the unmanned tractor usually tracked the reference curve generated by the planning algorithm with high accuracy and responsiveness. However, the replanning was performed to discard the remaining part of the previous solution, when the motion state deviated significantly from the target trajectory under external disturbances. The U-turn field experiment showed that the mean absolute error (MAE) and the root mean square error (RMSE) of speed decreased by 42.31%, and 50.75%, respectively, compared with the control group. The average absolute value and the variance of acceleration decreased by 8.26%, and 16.36%, respectively, while their jerk decreased by 7.65% and 14.23%, respectively. The variance of engine speed, torque percentage, instantaneous fuel consumption, and total fuel consumption decreased by 63.36%, 60.26%, 71.25%, and 2.37%, respectively. The straight-line navigation velocity adjustment experiment showed a similar optimization trend. The MAE and the RMSE of speed decreased by 9.45%, and 11.14%, while the average absolute value and variance of acceleration decreased by 6.03% and 13.68%, respectively. The average absolute value and variance of jerk decreased by 1.55% and 3.59%, respectively. The variance of engine speed decreased by 31.78%, the variance of engine torque percentage decreased by 25.13%, and the variance of instantaneous fuel consumption decreased by 31.82%. The total fuel consumption decreased by 2.48%. The stability of speed regulation was significantly improved to reduce fuel consumption during operation. The smooth control of velocity switching can fully meet the requirement for autonomous CVT tractors.

A perception system is one of the most important components for the autonomous driving of agricultural machinery. However, there are only a few perception datasets specifically designed for agricultural scenarios, due to their difference from the typical urban scenarios in previous studies. In contrast to the urban examples, the agricultural applications can suffer from harsh working circumstances. It is often required for the perception sensors and algorithms. In this study, a low-cost perception system was proposed for the two-stage detection using millimeter-wave radar and a monocular camera. 3D object detection was then performed on the autonomous driving of the agricultural machinery under agricultural scenarios. Firstly, a multimodal perception dataset of the agricultural scenes was constructed to incorporate the LiDAR (light detection and ranging), INS (inertial navigation system), camera, and millimeter-wave radar data with a hardware-level data synchronization and target-level data annotation. Then the middle fusion strategy was used to build a neural network model, known as CFPNet. Preliminary detection of the target was also implemented with the improved network of the center point detection. Furthermore, the radar point cloud features were extracted from the frustum region of interest to supplement the image features. Finally, the preliminary detection information and radar feature were combined to perform a secondary detection. The 3D object attributes (depth, direction, and velocity) were regressed concurrently. The results show that the mAP (mean average precision) of the CFPNet on the self-built multimodal dataset of the agricultural perception was 86.5%, which was 5.5 percentage points higher than the baseline, and the mATE (mean average translation error) was 0.197 m lower than the baseline. An additional experiment on small object detection was conducted to verify the effectiveness of the CFPNet. The better performance was achieved in a recall rate of 1 for the selected small objects, which was 0.3 higher than before the improvement, indicating the better performance of the detection. Deployment experiments were conducted to test the applicability of the CFPNet. A frame rate of 7.4 frames per second was 211% of the baseline in the low-computing agricultural scenarios. Experiments on the public datasets were conducted to test the CFPNet in the rest scenarios. The favorable performance was achieved on the NuScenes public dataset, with the mATE, mASE, and mAVE of 0.792 m, 0.236, and 0.52 m/s, respectively. Since the CFPNet was specifically designed for monocular cameras, its mAP lagged behind. Furthermore, the CFPNet can directly provide the speed information of the target without the preceding and following frames. This finding can provide a feasible solution and technical support for the 3D object detection in agricultural scenarios, especially with low computing power.

Agricultural machinery has been closely related to national food security during farming, sowing, and harvesting. The field harvested acreage of agricultural machinery can be required to accurately calculate the subsidy distribution, fee collection, efficiency, and income evaluation. It is also an urgent need for a precise and effective harvested acreage. This study aims to calculate the harvested acreage of wheat harvesters using spatiotemporal trajectories and grid key points. The calculation was also considered on the positional relationship between boundary grid key points and trajectory vector rectangles. The mathematical expectation of various occupancy scenarios was then proposed to effectively remove the overlapping areas for high accuracy. A grid key point algorithm was introduced to fully analyze the boundary grids, in response to the failure of the grid buffer zone. Probable situations were then determined at the boundary grids, according to the positional relationship between grid vertices and center points with vector rectangles. Mathematical expectation values were computed for each scenario during area calculation. The trajectory data of combined harvesters was obtained in the period of wheat harvest from the National Agricultural Machinery Big Data Platform. Among them, the parameters also included the latitude and longitude of the field, time, vehicle ID, speed, and direction. The trajectory data was selected as the meter-level positioning and the sampling frequency of 0.2 Hz. Data cleaning and segmentation were performed on the fields and roads. The test dataset of trajectory was captured from 50 farmland plots in the major wheat-producing regions (Hebei, Henan, and Shandong) for the years 2022 and 2023. A sensitivity analysis was conducted on the length of the bottom grid edge. Ultimately, the length of the bottom grid edge was determined to be one meter for both the grid buffer zone and the grid key point. A better balance was also obtained to consider the accuracy of area calculation and the speed of the algorithm. Three approaches (distance swath, grid buffer zone, and grid key point) were utilized to evaluate the operational width of the selected combine harvesters and the calculated area of the 50 farmland plots. The average error of calculation was then compared with the true values. Results showed that the grid key point shared an experimental average error of 2.66%, which was 69.45 and 1.25 percentage points lower than the distance swath and the grid buffer zone, respectively. The grid key point also solved the overlapping operation areas in farmlands. The higher accuracy of calculation was then achieved, compared with similar algorithms. Additionally, the versatility of the improved model was achieved by only requiringthe trajectory data and working width during rice harvesting, corn harvesting, or wheat sowing. In conclusion, the accuracy of calculation was effectively improved for the harvested acreage. The finding can also provide a strong reference for the subsidies, cost control, and budget planning in the assessment of agricultural production.

Autonomous agricultural machinery has been emerged as one of the core strategies in recent years. It is ever pressing from the aging population, shrinking agricultural workforce, large-scale rural migration, and agricultural labor costs in China. Autonomous agricultural machinery can be expected to alleviate the labor shortages for the high efficiency and precision. Fully functional unmanned farm can also enhance the agricultural productivity and sustainability. In this study, a comprehensive review was presented on the technical requirements of autonomous agricultural machinery. An emphasis was also put on the key research advancements in China, including the environmental perception, high-precision agricultural mapping, autonomous positioning and navigation, path planning, and tracking control. One of the hallmark features of unmanned farms was attributed to the seamless and fully automated transfer of machinery among storage facilities and fields. Specifically, the autonomous machinery was automatically departed from its designated storage location, thus navigating the roads to perform field operations, and finally returning to park precisely at its original position in the shed after task completion. As such, the autonomous machinery was used in fields, farm roads, and agricultural sheds, indicating the great potential to farming. Nevertheless, the unmanned farm system was remained underdeveloped within the storage shed. Machinery parking, implement attachment, and departure still heavily depended on the manual intervention. Full automation over all scenarios was required the advanced positioning and navigation. Precise perception to shed environment was also necessary for the safety and efficiency during operations. Storage sheds were often densely populated with machinery, rather than the open and relatively predictable environments of fields and farm roads. The maneuvering space was limited to the more rapid and accurate perception and navigation, compared with the open fields. While the significant progress was found in the autonomous machinery under the scenarios of fields and road. The perception and autonomous parking technologies were less developed for shed operations so far. These challenges were also addressed to bridge the automation gap, and then unlock the full potential of unmanned farms. Much effort was focused mainly on the automatic parking of autonomous agricultural machinery, particularly on returning to the shed after field operations. Indoor positioning and precise navigation were also explored with the perception during automatic parking. Some procedures were involved in the indoor navigation from the shed entrance to the parking spot without Global Navigation Satellite System (GNSS) signals. The safe and efficient routes were determined for the final docking after environmental perception and machinery position. The fixed- and non-fixed routes were categorized for the positioning and navigation. Fixed-route navigation offered the simplicity, precision and stability, including the visual navigation (lane tracking and visual marker localization), rail guidance, and magnetic navigation. Safe zones of shed were also delineated to enhance the accuracy of position and parking. In contrast, the non-fixed-route navigation was provided for the superior flexibility, scalability, and adaptability to the complex and dynamic indoor environments, such as the external source-based positioning (e.g., radio frequency identification, WiFi, bluetooth, and ultra-wideband) and simultaneous localization and mapping (SLAM). Finally, the technical and practical challenges were summarized for the autonomous agricultural machinery. Specifically, the diverse technologies were integrated for the dynamic environments, particularly for the robust safety mechanisms. Future directions were outlined to integrate the multiple technologies for all-weather operation in the complex environments. Autonomous agricultural machinery can be expected to serve as the critical technology in unmanned farming. In turn, the operational quality and efficiency can also be enhanced in the modern agriculture.