Agricultural robots are one of the important means to promote agricultural modernization and improve agricultural efficiency. With the development of artificial intelligence technology and the maturity of Internet of Things (IoT) technology, people put forward higher requirements for the intelligence of robots. Agricultural robots must have intelligent control functions in agricultural scenarios and be able to autonomously decide paths to complete agricultural tasks. In response to this requirement, this paper proposes a Residual-like Soft Actor Critic (R-SAC) algorithm for agricultural scenarios to realize safe obstacle avoidance and intelligent path planning of robots. In addition, in order to alleviate the time-consuming problem of exploration process of reinforcement learning, this paper proposes an offline expert experience pre-training method, which improves the training efficiency of reinforcement learning. Moreover, this paper optimizes the reward mechanism of the algorithm by using multi-step TD-error, which solves the probable dilemma during training. Experiments verify that our proposed method has stable performance in both static and dynamic obstacle environments, and is superior to other reinforcement learning algorithms. It is a stable and efficient path planning method and has visible application potential in agricultural robots.

/pdf/the-intelligent-path-planning-system-of-agricultural-robot-2od1e01y.pdf

The Intelligent Path Planning System of Agricultural Robot via Reinforcement Learning

Our work is focused on developing an autonomous robot to monitor greenhouses and large fields. This system is designed to operate autonomously to extract useful information from the plants based on precise GPS localization. The proposed robot is based on an RGB camera for plant detection and a multispectral camera for extracting the different special bands for processing, and an embedded architecture integrating a Nvidia Jetson Nano, which allows us to perform the required processing. Our system uses a multi-sensor fusion to manage two parts of the algorithm. Therefore, the proposed algorithm was partitioned on the CPU-GPU embedded architecture. This allows us to process each image in 1.94 s in a sequential implementation on the embedded architecture. The approach followed in our implementation is based on a Hardware/Software Co-Design study to propose an optimal implementation. The experiments were conducted on a tomato farm, and the system showed that we can process different images in real time. The parallel implementation allows to process each image in 36 ms allowing us to satisfy the real-time constraints based on 5 images/s. On a laptop, we have a total processing time of 604 ms for the sequential implementation and 9 ms for the parallel processing. In this context, we obtained an acceleration factor of 66 for the laptop and 54 for the embedded architecture. The energy consumption evaluation showed that the prototyped system consumes a power between 4 W and 8 W. For this raison, in our case, we opted a low-cost embedded architecture based on Nvidia Jetson Nano.

/pdf/mapping-agricultural-soil-in-greenhouse-using-an-autonomous-8wtta45o.pdf

Mapping Agricultural Soil in Greenhouse Using an Autonomous Low-Cost Robot and Precise Monitoring

Waste detection is one of the main problems preventing the realization of automated waste classification, which is a basic function for robotic arms. In addition to object identification in general image analysis, a waste-sorting robotic arm not only needs to identify a target object but also needs to accurately judge its placement angle so that it can determine an appropriate angle for grasping. In order to solve the problem of low-accuracy image detection caused by irregular placement angles, in this work, we propose an improved oriented waste detection method based on YOLOv5. By optimizing the detection head of the YOLOv5 model, this method can generate an oriented detection box for a waste object that is placed at any angle. Based on the proposed scheme, we further improved three aspects of the performance of YOLOv5 in the detection of waste objects: the angular loss function was derived based on dynamic smoothing to enhance the model’s angular prediction ability, the backbone network was optimized with enhanced shallow features and attention mechanisms, and the feature aggregation network was improved to enhance the effects of feature multi-scale fusion. The experimental results showed that the detection performance of the proposed method for waste targets was better than other deep learning methods. Its average accuracy and recall were 93.9% and 94.8%, respectively, which were 11.6% and 7.6% higher than those of the original network, respectively.

/pdf/improved-method-for-oriented-waste-detection-2vmpcap9.pdf

Improved Method for Oriented Waste Detection

Vespa velutina nigrithorax, a species of hornet, is spreading globally, with increasingly negative effects on human health. To effectively eliminate V. velutina, its nest should be destroyed and its queen removed; however, the nests are difficult to reach. Thus, we analyzed the requirements for a drone-assisted hornet exterminator using field observations and physical tests on a sample hornets’ nest, and based on these, a UAV exterminator equipped with a nest-perforating device (based on an airsoft rifle) and pesticide-spraying system was designed and manufactured. Pesticides and bullets were manufactured using ecofriendly materials. An actuator at the rear of the device adjusted the pitch of the perforator and sprayer, and a monitoring system was installed to aid the operator in targeting. The operating parameters of the UAV exterminator were evaluated in laboratory tests, with a spray distance of 5 m deemed suitable. To evaluate the system’s pest-control performance, several V. velutina nests were targeted in field tests. An insecticidal effect of over 99% was achieved using two pyrethrum-based pesticides (15% pyrethrum extract and 10% pyrethrum extract with additives). In addition, compared to conventional nest-removal methods, the UAV exterminator reduced the work time by 85% and the cost by 54.9%.

/pdf/exterminator-for-the-nests-of-vespa-velutina-nigrithorax-1np3kynl.pdf

Exterminator for the Nests of Vespa velutina nigrithorax Using an Unmanned Aerial Vehicle

Research on site selection of agricultural internet of things nodes based on rapid terrain sampling

High-precision position estimations of agricultural mobile robots (AMRs) are crucial for implementing control instructions. Although the global navigation satellite system (GNSS) and real-time kinematic GNSS (RTK-GNSS) provide high-precision positioning, the AMR accuracy decreases when the signals interfere with buildings or trees. An improved position estimation algorithm based on multisensor fusion and autoencoder neural network is proposed. The multisensor, RTK-GNSS, inertial-measurement-unit, and dual-rotary-encoder data are fused with Extended Kalman filter (EKF). To optimize the EKF noise matrix, the autoencoder and radial basis function (ARBF) neural network was used for modeling the state equation noise and EKF measurement equation. A multisensor AMR test platform was constructed for static experiments to estimate the circular error probability and twice-the-distance root-mean-squared criteria. Dynamic experiments were conducted on road, grass, and field environments. To validate the robustness of the proposed algorithm, abnormal working conditions of the sensors were tested on the road. The results showed that the positioning estimation accuracy was improved compared to the RTK-GNSS in all three environments. When the RTK-GNSS signal experienced interference or rotary encoders failed, the system could still improve the position estimation accuracy. The proposed system and optimization algorithm are thus significant for improving AMR position prediction performance.

/pdf/improved-position-estimation-algorithm-of-agricultural-cfm56mfz.pdf

Improved Position Estimation Algorithm of Agricultural Mobile Robots Based on Multisensor Fusion and Autoencoder Neural Network

Dynamic Beehive Detection and Tracking System Based on YOLO V5 and Unmanned Aerial Vehicle

Introduction Canopy stomatal conductance (Sc) indicates the strength of photosynthesis and transpiration of plants. In addition, Sc is a physiological indicator that is widely employed to detect crop water stress. Unfortunately, existing methods for measuring canopy Sc are time-consuming, laborious, and poorly representative. Methods To solve these problems, in this study, we combined multispectral vegetation index (VI) and texture features to predict the Sc values and used citrus trees in the fruit growth period as the research object. To achieve this, VI and texture feature data of the experimental area were obtained using a multispectral camera. The H (Hue), S (Saturation) and V (Value) segmentation algorithm and the determined threshold of VI were used to obtain the canopy area images, and the accuracy of the extraction results was evaluated. Subsequently, the gray level co-occurrence matrix (GLCM) was used to calculate the eight texture features of the image, and then the full subset filter was used to obtain the sensitive image texture features and VI. Support vector regression, random forest regression, and k-nearest neighbor regression (KNR) Sc prediction models were constructed, which were based on single and combined variables. Results The analysis revealed the following: 1) the accuracy of the HSV segmentation algorithm was the highest, achieving more than 80%. The accuracy of the VI threshold algorithm using excess green was approximately 80%, which achieved accurate segmentation. 2) The citrus tree photosynthetic parameters were all affected by different water supply treatments. The greater the degree of water stress, the lower the net photosynthetic rate (Pn), transpiration rate (Tr), and Sc of the leaves. 3) In the three Sc prediction models, The KNR model, which was constructed by combining image texture features and VI had the optimum prediction effect (training set: R2 = 0.91076, RMSE = 0.00070; validation set; R2 = 0.77937, RMSE = 0.00165). Compared with the KNR model, which was only based on VI or image texture features, the R2 of the validation set of the KNR model based on combined variables was improved respectively by 6.97% and 28.42%. Discussion This study provides a reference for large-scale remote sensing monitoring of citrus Sc by multispectral technology. Moreover, it can be used to monitor the dynamic changes of Sc and provide a new technique for gaining a better understanding of the growth status and water stress of citrus crops.

/pdf/estimating-stomatal-conductance-of-citrus-under-water-stress-2c1rnxb8.pdf

Gaotian Liang

Papers

Improved Position Estimation Algorithm of Agricultural Mobile Robots Based on Multisensor Fusion and Autoencoder Neural Network

Dynamic Beehive Detection and Tracking System Based on YOLO V5 and Unmanned Aerial Vehicle

Research on site selection of agricultural internet of things nodes based on rapid terrain sampling

Estimating stomatal conductance of citrus under water stress based on multispectral imagery and machine learning methods