Real- time UAV weed scout for selective weed control by adaptive robust control and machine learning algorithm

Abstract. Typically, herbicides are applied uniformly to the entire field rather than be applied to the critical areas where they‘re most needed. And existing weed identification approaches, such as manual visual scouting and ground-based proximal sensing, are labor intensive, time consuming, and do not scale easily to large areas. To facilitate more efficient farming, new technologies must be developed that can achieve optimal weed management and increase the crop yield by a magnitude. To date, rapid advances in UAV technology has enabled sufficient lifting capacity and endurance to support high resolution imagery over large areas. With the FAA planning to integrate UAVs into the national airspace (NAS), UAV-based aerial weed scouting is now possible to bring promising significant benefits for farmers. An unmanned aerial vehicle based aerial weed scout with machine vision system was proposed and developed for the weed identification with the sensing capability from far above the field to close to the canopy, and measuring the plant/weed density (weed infestation rate)/weed species. The canopy was segmented from the background by greenness selection and thresholding. The crop row and its centroid line was calculated and masked by its pixel density. The anomalous weed patches between the crop rows was identified and its population was mapped. For species identification, a machine learning approach is used. Convolutional Neural Network (CNN) as one of deep learning method has the advantage of less neuron weight, easy for training and good robust performance in displacement, scale and deformation, which is promising for the weed specifies classification and probability assessment. The previous weed distribution mapping and individual weed extraction helps to generate the training data and highlights their own features, which could ensure the weed identification classifier robustness and efficiency. The preliminary result shows the specific weed type could be classified probably. The results show that the proposed approach will provide an efficient and promising tool for farmers‘ weed identification and the resulting infestation mapping will enable a better weed control recommendation.