UC Davis

Road facilitates invasive plant species' early establishment and spread, which can impact ecosystem services and cause economic loss. Monitoring invasive species populations along the road is important in roadside integrated vegetation management (IVM). A roadside weedy and invasive species map can assist in developing species population models and designing proper weed management strategies. However, the traditional road survey requires unrealistic human labor and time to map invasive species at large scales. A novel weed mapping system was developed to retrieve species location data by integrating Google Street View (GSV) imagery and object detection based on deep learning algorithms. The target species of this feasibility study was johnsongrass (Sorghum halepense). We trained the detection network, You Only Look Once (YOLOv2), with 911 johnsongrass roadside images retrieved from Google Street View. YOLOv2 is a fast and accurate deep neural network. The trained detection model could detect johnsongrass in GSV images, and output bounding boxes with the target species' confidence scores. The trained model was then applied to a large image dataset of ~270,000 images along 135,000 km of roads in California, Oregon, Washington, and Nevada. The network detected 2,031 new johnsongrass records along roads in these four states, and the location of each image was used to create a map of the johnsongrass population. Our current deep learning model has 85% recall, 73.9% precision, and 77.5% overall accuracy on the testing dataset, which included 2,040 images. The model also has a 30% false positive rate (FPR). Work is in progress to reduce the FPR. Using our novel AI-based method, the estimated cost of the weed survey in four states is $3,570, while the traditional road surveys with cars at the same scale will cost at least $63,558 without considering associated risks, such as car accidents. Besides that, traditional road surveys require six months, but the automated weed survey only requires a few days with a trained detection network. The automated mapping scheme can apply to other weedy and invasive species, and it is possible to map this weed (and others) on a much larger scale, which is the focus of our future work.