Research On UAV-based Monitoring Methods For Citrus Under Complex Environments

Agricultural production is one of the foundations of social stability and prosperity,so agricultural technology has always been the focus of research.When the general technology is mature,it will often be applied to the field of agricultural production,especially UAV technology and DNN technology is more popular recently.When the two technologies are combined,to a certain extent,the difficult problems of automatic high-altitude monitoring can be solved.At the same time,there are many difficulties in the field work of agricultural production,which need to overcome the complex factors of dangerous areas.The emergence of UAV automatic identification greatly alleviates such problems.Recently,due to the continuous exploration of deep neural network,many successful DNN research works have emerged.While solving various common tasks,it also plays a great role in different vertical fields.Although the existing algorithm research is far from mature and complete,and has some limitations,its effectiveness has been verified for a long time.Therefore,in this paper,according to the different flight altitude of UAV,using DNN training model,the full supervision and weak supervision methods based on deep learning are discussed in depth.The related improved algorithm is used to complete different tasks,and the overall scheme of citrus monitoring is proposed,including high-altitude scene classification,low-altitude citrus plant detection and short-range citrus fruit counting,which overcomes the limitations of related work.The completed work is as follows:(1)The problem of orange planting area identification by high altitude UAV is studied,and a scene classification algorithm based on UAV high altitude monitoring is proposed.Compared with the traditional algorithm work,the full supervised algorithm in this paper improves the attention module to train the model.On the obtained citrus monitoring data set,the high-altitude citrus scene classification algorithm AMDRC-Net achieves very competitive test results.It supports the UAV to effectively identify the citrus planting area,and provides help for the subsequent intelligent monitoring of citrus growth.(2)This paper studies the problem of orange plant recognition by low-altitude UAV,and proposes AF-AMDRC structure based on anchor free detection framework.Based on the low altitude monitoring data of citrus,the AMDRC-Net of high-altitude scene classification is used as the backbone network.Considering that the target detection model consumes a lot of computation and has practical problems,an improved anchor free detection framework is proposed to solve these problems.At the same time,the classical gradient descent algorithm is analyzed.Finally,the AF-AMDRC structure is verified to make the intelligent monitoring more efficient and reliable.Compared with the traditional algorithm,the performance of the proposed framework is better on the low altitude monitoring data set.(3)The problem of citrus fruit counting in close range UAV is studied,and a fully supervised citrus close range detection algorithm based on two stages is proposed.Compared with the traditional two-stage method,the proposed Fast A-R-CNN structure well balances the performance and efficiency of the two-stage detector,and optimizes the citrus data set,model structure and stability.At the same time,by introducing the classical detection structure,a lot of theoretical knowledge and task difficulties are elaborated.For the intelligent recognition of citrus fruit by UAV monitoring,it plays an indispensable role in the total monitoring scheme,which can obtain the pre-production value of planting land in real time and help managers to make reasonable sales scheme.(4)The problem of citrus fruit counting in UAV complex environment was studied,and a method of citrus fruit detection based on weak supervision was proposed.Finally,the experiment deals with the complex environment and weak labeled data.The results show that the A-PSOL method is better.The algorithm based on weak supervision is of great significance to the overall monitoring scheme,which not only greatly reduces the cost of making sample labels,but also enhances the detection ability of the model.In the face of the complexity of the actual demand,the weak supervision method makes the monitoring model more stable and reliable,and it is accurate to know the growth status of citrus.