Weed location and recognition based on UAV imaging and deep learning

Abstract: To locate and identify weeds in a wheat field efficiently, an unmanned aerial vehicle (UAV) based imaging method was developed in this study.  A weed detection model based on image data through deep learning was developed and implemented.  The model uses the YOLOV3-tiny network to detect the pixel coordinates of weeds in images.  It acquires the position of weeds by converting the pixel coordinates to the geodetic coordinates.  The identified weeds were marked on the prescription map.  The algorithm was implemented and tested using a commercial DJI Phantom 3 UAV.  This study tested the performance of YOLOV3 and YOLOV3-tiny and found that YOLOV3-tiny was more suitable for mobile devices.  The performance of YOLOV3-tiny at different thresholds was tested.  The test results show that the model performs optimally when the threshold of the YOLOV3-tiny network is 0.5, under this condition, the mean Average Precision (mAP) is 72.5%, the Intersection-over-Union (IOU) is 80.12%, and the mobile device processing speed is 2FPS.  After testing and analyzing weed positioning, results show the average positioning error is10.31 cm, which is extremely small in agricultural operations.  The UAV-based weed position detection system can locate and identify weeds in the crop field at a high speed, efficiently and effectively. Keywords: UAV, deep learning, weed location, weed recognition, imaging method, target detection, Android APP DOI:  10.33440/j.ijpaa.20200301.63 Citation: Zhang R F, Wang C, Hu X P, Liu Y X, Chen S, Su B F. Weed location and recognition based on UAV imaging and deep learning. Int J Precis Agric Aviat, 2020; 3(1): 23–29.