Abstract:
Rice is a crop feeding more than 3 billion people across the world. Continuous improvements in farming techniques are helpful to farmers looking for high yields. This thesis reports on the development of a system to help farmers monitor a rice crop using deep learning techniques and UAVs equipped with vision sensors. The system uses UAV-collected data and state of the art deep learning methods to estimate plant counts, detect weeds in early stages of growth, and monitor plant health.
A test rice field of size 2000m2 is chosen for the study. The fields monitored with UAVs from low altitude in the 5m to 20m range over the lifetime of the crop. The data collected include videos captured from two cameras (RGB and NIR). Open Drone Map (ODM), which uses Structure from Motion (SfM) techniques, is used to generate ortho-images of the field. A semi-manual method is used to stitch images in cases of feature match failure. The system uses CSRNet to estimate plant count and generate density maps from the images collected from 3m to 6m above the ground in the early stages of growth. The system uses YOLOv3 to detect weeds during the growing stages of the rice. The system shows good performance in plant counting, with MAE of 27.94, and good results in detecting weeds, with a mAP of 71% on UAV-collected data. The system also uses a low-cost modified RGB camera to capture data for NDVI image collection.