In the master's thesis, we address the problem of recognizing and counting corn plants using deep learning methods for precision agriculture. Based on images of cornfields captured by a drone, a dataset with labeled leaves and plants was created to train the deep learning model BranchedERFNet. This model, which enables simultaneous semantic and instance segmentation of leaves and plants, was trained and tested across different phenological stages and image capture conditions. The trained model successfully detects and counts corn plants with an accuracy of 82\%, despite some limitations in segmentation quality. The performance is higher for corn plants in earlier phenological stages. The results indicate that the model performs well at different image capture heights but struggles when leaves of neighboring plants are densely intertwined. The accuracy of segmentation at different phenological stages of the plants was 25\% and is a key factor in the less successful precise determination of the centers of corn plants.
|
