New methods to improve detection and classification on waterfowl and tree in aerial images

Monitoring waterfowl populations and their classification is vital for wetland conservation, and deep learning has recently demonstrated significant potential in detecting waterfowl from aerial images. This dissertation introduces several innovative methods to enhance waterfowl detection and classification. First, the Self-Rectification Network (SRN) was developed to improve detection performance in complex backgrounds by reducing false positives. SRN employs a novel sampling mechanism that ensures a balanced ratio of positive and negative proposals in each training batch and prioritizes challenging cases, achieving a 10.1 percent improvement on hard-to-detect images. Second, a Size-Related Non-Maximum Suppression (NMS) technique was introduced to eliminate false detections of objects with abnormal sizes, leveraging the consistent pixel dimensions of waterfowl in aerial imagery. This approach improved detection accuracy by approximately 3 percent on specific datasets. Finally, to address the labeling burden, a class-balanced sampling strategy was devised, enabling effective training with 50 percent labeled data and significantly reducing the time and effort required for annotation. Tree canopy classification, another focus of this dissertation, holds great po- tential for monitoring environmental diversity and stability. While trees can be categorized by family, genus, and species, existing classifiers often overlook the hi- erarchical relationships among these categories. To address this gap, ConsistentNet was developed to ensure consistency in predictions across these hierarchical levels. A novel consistency loss function supervises these relationships during training, resulting in superior performance on datasets with class-subclass structures. Consis- tentNet demonstrated a 5 percent improvement in accuracy over baseline models on tree canopy classification datasets and a 2 percent improvement on the CIFAR-100 dataset. Together, these advancements underscore the promise of deep learning for ecolog- ical monitoring and classification tasks, offering robust solutions for challenging scenarios and hierarchical datasets.

Name thesis:degree_name

Ph. D.

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Computer Science (MU)

Grantor dc:publisher

University of Missouri--Columbia

Year dc:date.issued

2024

Author dc:creator

• Zhang, Yang

Advisor dc:contributor.advisor

• Shang, Yi

Language dc:language.iso

eng, English

DOI

https://doi.org/10.32469/10355/107644

OAI identifier oai:identifier

oai:mospace.umsystem.edu:10355/107644