Developing an Algorithm for Robotic Precision Application of Crop Protection Products

Abstract

The existing range of plant identification methods and tools is considered limited in real agrotechnical tasks. The image parameters tend to differ significantly in applied solutions. (Research purpose) To develop an algorithm for crop plant recognition by a robotic device using a state-of-the-art convolutional neural network (R-CNN) and deep learning technology. (Materials and methods) A robotic device has been developed for variable rate application of plant protection products able to recognize both useful crops and weeds, determine the area of processing, namely the coordinates of the processing center and the processing radius. Mask R-CNN and Deeplabv3 plus segmenting neural networks were chosen for crop (white head cabbage) detection. The network-based algorithm detects, segments, and positions plants based on a dataset collected in the image-mask and COCO dataset formats. The data set was formed by aerial photography using an unmanned aircraft. The original images are taken by Xiaovv HD Web USB 150 degree Full HD 1080P webcam and Logitech C270 HD 720p webcam. The trained neural network for the robotic device was installed on the Nvidia Jetson AGX Xavier platform. (Results and discussion) As a result of assessing the accuracy of the model on the test data, the following values were obtained: the number of plants detected is 98 percent, the accuracy of contour detection is 94 percent. (Conclusions) It is proved that the trained neural network can be applied to any cultivated crops, taking into account the heterogeneity of their location in the field, soil types, and the percentage of weeds. As a result, the model is trained to extract the bounding box coordinates and the object (cabbage) location by pixels with the required accuracy for both synthetic and real data.