Accurate prediction of regional water evapotranspiration can greatly contribute to the rational utilization of regional water resources for the water resources saving. Crop evapotranspiration can be one of the most important indicators for the water evapotranspiration status of crops, in order to evaluate the soil water balance of farmland and the water management of farmland. However, the calculation of crop evapotranspiration requires a large amount of meteorological factor data. There is often redundant data with the low correlation with the evapotranspiration in meteorological data, which seriously affects the prediction accuracy and efficiency of the model. It is a high demand to extract the core environmental factors for the less-factor prediction model. The classification and regression tree (CART) can be used to make feasible for the large data sources in a short time. Then, Less factors can be extracted for the predictive analysis in a more scientific way. In this study, a less-factor water evapotranspiration model was constructed to efficiently and accurately predict the evapotranspiration data. The core factors were selected from multiple meteorological factors. 23 typical stations were selected in the nine agricultural regions, and then collected data on eight meteorological factors, such as precipitation, and sunshine hours. The meteorological factors were ranked in the order of importance using CART. The top 3-5 meteorological factors were then selected to predict the evapotranspiration using the extreme learning machine (ELM) model. At the same time, the ELM model was optimized using genetic algorithm (GA), particle swarm optimization (PSO), sparrow search algorithm (SSA). Three optimization algorithms (GA-ELM, PSO-ELM, SSA-ELM) were used to construct a less-factor hybrid optimization water evapotranspiration prediction model. Relative root mean square error (RMSE), coefficient of determination (R²), mean absolute error (MAE), and Nash-Sutcliffe coefficient (NSE) were used to evaluate the performance of the ELM and optimization model. The results showed that: 1) The main influencing factors were ranked in the order of the precipitation, sunshine duration, mean station pressure, daily maximum temperature, and average relative humidity using the importance ranking of the CART algorithm. 2) The PSO-ELM model presented the highest prediction accuracy among the three optimization algorithms. Specifically, the RMSE of evapotranspiration prediction for the 23 stations was ranged from 6.608 to 22.077 mm/d, the NSE of 0.824-0.998, R² of 0.908-0.995, and MAE of 5.075-16.677 mm/d. In addition, the RF and SVR model with the strong generalization performance were selected to compare with the PSO-ELM model. The prediction performance of three models was slightly improved with the increase of the input factors, indicating the slight overall improvement. The three models shared the strong generalization and robustness. The meteorological factors were input into the prediction model, according to the order of importance. The meteorological factors ranked the 4th and 5th were relatively less important, where the prediction accuracy was slightly improved with the increase of the input parameters. The PSO-ELM model presented the highest prediction accuracy among the three models. 3) The ELM model performed better applicability in the Yunnan-Guizhou Plateau region, the Sichuan Basin, and the surrounding areas. The three optimization algorithms showed the better applicability in the South China and the Yunnan-Guizhou Plateau region, with the highest applicability of the PSO-ELM model. The findings can provide an important reference for the crop water demand calculation in nine major agricultural regions in China.

The development of intelligent decision-making technologies for agricultural machinery has been a critical research focus over the past decade, as they serve as the key enabler for autonomous operation and large-scale cooperative tasks in modern farming. The primary goal of previous studies was to improve operational efficiency, adaptability, and coordination of agricultural equipment under dynamic and uncertain field conditions. While many existing works concentrated on specific stages of the integrated “perception–control–execution” process, comprehensive reviews of decision-making architectures and their systematic frameworks remained limited. Against this background, the present study aimed to systematically examine the state of research on intelligent decision-making in agricultural machinery, with particular attention to perception modeling, decision-making strategies, and adaptive control mechanisms. To achieve this objective, the study first adopted a comparative approach to survey multi-source heterogeneous information perception techniques. The methods analyzed included state estimation models, semantic fusion strategies, and scene-adaptive configurations of sensors and equipment. Special emphasis was placed on multi-modal fusion structures that integrated environmental information with operational states, enabling robust perception under varying task requirements. This phase of the review also considered the challenges of processing noisy data, ensuring sensor reliability, and establishing consistent standards for agricultural environments that are inherently unstructured and variable. Subsequently, the review examined three distinct categories of intelligent decision-making generation methods: rule-driven approaches, optimization-driven models, and learning-driven algorithms. Each category was assessed according to its applicable conditions, computational requirements, and adaptability to real-world agricultural environments. Rule-based methods emphasized interpretability and transparency but lacked flexibility when conditions deviated from predefined patterns. Optimization-driven methods, such as task scheduling and path planning algorithms, were particularly effective for resource allocation and route efficiency, though they often demanded extensive computational power. Learning-driven methods, most notably reinforcement learning and deep learning approaches, offered superior adaptability and the capacity to evolve through experience, but they faced limitations in terms of training data requirements, stability, and deployment feasibility in real time. The results of this comprehensive review demonstrated that multi-modal perception modeling successfully improved the robustness of environmental sensing, particularly under diverse and uncertain field conditions. Rule-driven decision-making approaches provided reliability and interpretability but exhibited limited flexibility in rapidly changing environments. Optimization-driven strategies showed strong performance in path planning and task allocation but required significant computational resources. Learning-driven methods, particularly those based on deep reinforcement learning, achieved superior adaptability and self-improvement, though their real-time applicability remained constrained by model complexity and data requirements. In terms of adaptive control, experimental studies highlighted the effectiveness of feedback-based real-time control mechanisms for adjusting machinery behaviors, thereby ensuring stability and safety during operation. Additionally, system-level intelligent decision-making frameworks were found to enhance multi-machine collaboration, farm management efficiency, and integration with intelligent management platforms. Evaluations of practical applications in crop monitoring, path optimization, collaborative scheduling, and smart farm platforms indicated promising results in improving both productivity and resource efficiency. In conclusion, intelligent decision-making technology for agricultural machinery is now positioned as a core driver of smart agriculture. However, challenges persist in perception robustness, adaptability of decision algorithms, and the maturity of large-scale applications. Current findings suggest that future progress depends on strengthening multi-modal and multi-factor coupling mechanisms, developing large-model-driven decision-making frameworks, and incorporating digital twin simulations for real-time optimization and validation. Moreover, progress in this field will require closer integration of artificial intelligence with domain-specific agricultural expertise, as well as cross-disciplinary cooperation among agronomy, computer science, and engineering. These advances are expected to unify agricultural efficiency, economic profitability, and ecological sustainability, thereby providing a strategic reference for the future development of smart agriculture and intelligent agricultural equipment.

To enhance the precision and efficiency of young apple fruit recognition in agricultural settings, where timely identification is crucial for effective crop management and yield optimization, this paper introduces an innovative approach utilizing the YOLOv8n model, bolstered by transfer learning techniques. The recognition of young apple fruits presents unique challenges due to the variability in size, shape, and maturity stages, as well as the dynamic environmental conditions within orchards, such as varying lighting and occlusion by leaves or branches. To tackle these challenges, the study embarked on constructing a comprehensive and diverse dataset that captures the essence of young apple fruits under a wide range of scenarios. High-resolution digital cameras were employed to gather images of young apple fruits at different times of the day and throughout various stages of growth. Special attention was given to incorporating a spectrum of lighting conditions, from bright sunlight to shaded areas, to ensure the model's robustness against illumination changes. Furthermore, fruits of different sizes and maturity levels were included to reflect the natural heterogeneity within an orchard. These images underwent rigorous preprocessing, which involved meticulous cropping to focus solely on the fruits and augmentation techniques like rotation, flipping, and color adjustment to artificially diversify the dataset and prevent overfitting. The study then compared the efficacy of the transfer learning approach against the conventional method of random weight initialization across several state-of-the-art object detection models, including RetinaNet, EfficientDet, YOLOv5n, YOLOv8n, and YOLOv10n. The objective was to assess which model, when equipped with pre-trained weights from a related domain, could best adapt to the task of identifying young apple fruits with high accuracy and efficiency. Notably, the YOLOv8n model, known for its balance between performance and computational efficiency, emerged as the top performer when enhanced with transfer learning. It achieved remarkable detection precision of 99.3%, indicating a high degree of correctness in identifying young apple fruits, coupled with a recall rate of 94.2%, which underscores its ability to capture most of the relevant instances in the dataset. The average precision (AP) of 97.0% and extended average precision (AP) across multiple IoU thresholds (83.1%) further solidified its superiority in this specific detection task.To push the boundaries of detection accuracy even further, this research introduced the EMCA (Efficient Multiscale Channel Attention) module, integrating it into the YOLOv8n framework to create the YOLOv8n-EMCA model. The EMCA module is designed to enhance the model's capability to extract and process multi-scale features, enabling it to attend more effectively to critical details across different spatial resolutions. This refinement led to slight improvements in precision (99.6%) and a notable jump in recall to 95.6%, indicating fewer missed detections. Additionally, the mean Average Precision at 50% IoU (mAP50) reached 97.3%, and the more stringent mAP50-95 metric improved to 88.2%, demonstrating the model's robustness across a range of IoU thresholds. This research not only contributes a novel methodology tailored for the identification of young apple fruits but also serves as a valuable reference for the development of similar systems for other fruit trees. The findings underscore the transformative potential of transfer learning and advanced attention mechanisms in bolstering object detection capabilities, with profound implications for advancing agricultural automation and machine vision applications. By leveraging these techniques, the agriculture sector can move closer to achieving precision farming, where real-time, accurate monitoring and decision-making become the norm.

Intelligent and automation technologies have been widely applied to agricultural production in recent years. A green apricot is one of the most significant economic fruits in Asian areas. However, traditional identification cannot fully meet large-scale production under complex environmental conditions in precision management. Fruit recognition and detection technologies can be required to enhance the accuracy and efficiency of large-scale production at present. In this study, an efficient and lightweight target recognition model (MCBPnet) was developed to automatically detect the greet apricot fruits using advanced deep learning. The CBAM (convolutional block attention module) was first integrated into the MobileNetV3 architecture. The precision of the model was significantly enhanced to identify the apricot fruits with the most rich image areas. The regions of an image were then prioritized with the most relevant to fruit detection. More computational resources were effectively allocated to the features closely associated with the task. As a result, the model structure, IRCBAM (inverted residual convolutional block attention module) was then developed to incorporate as the backbone network of MCBPnet. BiFPN (Bi-directional feature pyramid network) was also introduced into the neck network, in order to further enhance the performance of the model. A feature pyramid was constructed to capture the multi-scale features from images using the BiFPN model. The fruits with varying sizes were then detected within a single frame. A high detection rate was also maintained in a wide range of fruit sizes and orientations. Moreover, some occlusions were also removed during detection, where the fruits were partially obscured by leaves or other fruits. The PConv (partial convolution) structure was utilized in the detection head. The high accuracy of detection was achieved, even when only a portion of a fruit was visible. As such, reliable recognition was obtained to increase its accuracy and robustness under conditions of partial visibility. The results demonstrated that the MCBPnet model achieved significant advantages across key performance metrics, including precision and mAP₅₀. Specifically, the high accuracy of detection was obtained with a precision of 0.988, an mAP₅₀ of 0.994. Furthermore, the detection speed of the MCBPnet model was 109.890 frames/s, which was 70.33% higher than that of the YOLOv8n model, and the model arithmetic was 6.1 G, which was 75.31% of that of the original one. This lightweight model was also deployed on edge devices, such as drones or handheld scanners, without any energy-intensive computational resources. In summary, the high-performance metrics were achieved by the MCBPnet model. The effectiveness of the integrated techniques was validated to maintain high accuracy during agricultural automation. Various indicators represented the significant advancement in the MCBPnet model. Advanced deep learning techniques and lightweight can be expected to improve the efficiency and accuracy of apricot detection during harvesting. The findings can provide technological support to the automated recognition and picking of green apricots. This MCBPnet model can greatly contribute to modern, intelligent, and sustainable crop management, thus driving the transition toward more efficient and intelligent agricultural practices.