Weed Identification And Location Based On Low Altitude Remote Sensing

Farmland weeds are one of the main factors affecting the growth of crops,farmland weed detection and distribution density mapping are of great significance for the accurate control of farmland weeds.To achieve the transformation of large-scale spraying of farmland to precision spraying,the distribution,species and density of weeds in a field need to be detected quickly and efficiently.The high maneuverability and flexibility of drones combined with deep learning algorithms have become an emerging method for the detection of farmland weed distribution density.This paper designs and develops a detection system for making farmland weed identification and distribution density mapping based on DJI drone platform and Android mobile devices.The system identifies farmland weeds based on the deep learning model,and determines the spatial position of the weeds by the transformation of the geodetic coordinate system.Finally,image data display,generation of weed distribution density map and other methods for integration testing were conducted by android platform.The main research work and results of this paper are as follows:(1)Optimizing the YOLOv3-Tiny weed target detection algorithm.Comparing the advantages and disadvantages of target detection algorithms such as Faster R-CNN,YOLO,SSD,etc.to weed target detection,YOLOv3-Tiny was used as the basic network,and the residual block is used to replace the YOLOv3-Tiny backbone network with residual blocks.Weed target detection models were obtained from two types of weeds,grasses and broadleaves.The model with the highest m AP was selected using F-Measure as the evaluation standard to evaluate the accuracy,recall,IOU,etc.,and the optimal model was selected.The final model has a m AP of 0.83 and an IOU of 0.751.(2)Weed spatial positioning and weed distribution density mapping.Decode the drone image data in Android mobile devices by calling FFmpeg and Media Codec to obtain the decoded image.Use the Darknet deep learning framework embedded in the Android device to perform weed detection on the image to obtain the type,quantity and position.By doing this,the position of the weed is the pixel coordinate position,then the conversion method of the pixel coordinate system,the world coordinate system,the conversion method of the WGS-84 coordinate system and the GCJ-02 coordinate system were used.First the position of the UAV in the WGS-84 coordinate system was transferred to the GCJ-02 coordinate system.Then the pixel coordinates of the weeds were transferred to the world coordinate system.Finally,the weed distribution density map was generated according to the detected weed location,type and quantity.(3)Integration and testing of UAV weed identification and positioning system.This research took DJI UAV and Android mobile devices as hardware development platforms.Based on the electronic map provided by Gaode and the MOBILE SDK software development support package provided by DJI,weed target detection algorithm and weed space location positioning in Android Studio Transplantation of the algorithm,weed distribution density map generation method,drone flight control method and Darknet deep learning framework were integrated in Android Studio.Finally,a comprehensive test and experiment were conducted on the UAV weed identification and positioning system.The test included system compatibility,stability,drone flight control function,drone image transmission function,weed target detection function and software operating performance.The results showed that the UAV weed identification and positioning system meets the design requirements.When the UAV is flying,the real-time detection speed is 1 FPS.The accuracy of the generated weed distribution density maps is 75.8%,78.6%,and 82.4%.