New methods to improve semantic segmentation on habitat classification
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
This dissertation develops new methods and a practical end-to-end system for habitat mapping and waterfowl monitoring from UAV imagery. We construct dedicated UAV datasets for wetlands and rivers, and propose a modular pipeline that combines deep neural segmentation backbones with the Segment Anything Model (SAM) and domain knowledge rules (P-KESS). we (i) train CNN backbones (FCN/UNet/PSPNet/DeepLabV3) for pixelwise segmentation, (ii) use the Segment Anything Model (SAM) to obtain fine object boundaries and assign labels via majority voting, and (iii) enforce P-KESS (Prior-Knowledge Enforced Semantic Segmentation), a set of domain prior knowledge rules learned from train dataset. We curate large UAV orthomosaics from five Missouri rivers and wetlands and evaluate against strong baselines. On river habitats, the two-stage design--dense segmentation followed by SAM-guided consolidation and P-KESS correction--achieves about 99% overall accuracy and about 95–96% average F1 (e.g., FCN+P-KESS: 99.3% overall accuracy, 95.5% , Avg F1 / 92.3% mIoU), while an independent assessment of SAM's raw boundary fidelity shows 98.6% mIoU and 99.1% pixel accuracy on high-contrast regions. Beyond our domain, on the public Semantic Riverscapes UAV benchmark (400 images, 14 classes), P-KESS consistently lifts all backbones over their bases (e.g., FCN+P-KESS: 87.1% OA, 82.2% Avg F, 79.8% mIoU), validating cross-dataset robustness. Finally, we design fast overlap-region detectors--including a bird-location-based method--to avoid double counting across sequential frames, and integrate the full workflow with automated reporting. Our overlap detection method maintained an error rate around 10% with faster performance compared to traditional techniques. The approach provides state-of-the-art accuracy, transparent corrections grounded in ecological priors, and scalable end-to-end analytics for conservation agencies across UAV platforms.
Table of Contents
PubMed ID
Degree
Ph. D.