Semisupervized Object Detection in UAV Images

Resource Type
Authors
Michael Fromm
Resource Date:
2018

This thesis investigates the effectiveness of machine learning algorithms for automatic detection of coniferous seedling data along Boreal seismic lines. In order to obtain a survival assessment and survey of the restoration process along these seismic lines, ground crews must undertake expeditions which are expensive, potentially hazardous and difficult to scale. Since the seismic lines cover a length of more than 10,000 km, an automated solution is necessary. The literature describes several machine learning applications using satellite data to extract information on topics such as forestation and expanse of deserts. In contrast, we conduct experiments on small sites on the basis of drone imagery. To this end we use algorithms from computer vision and apply them to the drone image data. We use convolutional neural networks as a feature extractor on the images. Subsequently, we train an object detector to spot the seedlings and annotate them with bounding boxes. In this work we evaluate the accuracy of modern object detectors such as Faster R-CNN with regard to remote sensing capacity of conifer seedlings. We further investigate the problem by doing several experiments which focus on the special environmental variables in nature, including seasons and flight height of the drone. These are necessary to understand what conditions are beneficial to support the machine learning process. Modern convolutional object detectors require huge amounts of data. Meanwhile, we conduct experiments on the amount of data needed to achieve high accuracy and we also investigate the influence of pretrained networks on the object detector. We further employ error analysis to understand how the object detector performs depending on the seedling size, to determine where further improvements are possible. Finally, we suggest further research options based on our findings.