Semester Award Granted

Spring 2025

Submission Date

May 2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Thesis/Dissertation Advisor [Chair]

Caiyun Zhang

Abstract

Wetlands play a significant role in the world’s hydrology, climate, and biodiversity. Even with the benefits and values wetlands provide to the environment, they have been undergoing loss and degradation due to natural and anthropogenic processes. To protect wetlands from loss and degradation and to restore their function, it is essential to develop a sustainable wetland monitoring system. One of the key elements of a wetland monitoring system is wetland mapping. This dissertation research developed an object-based deep learning protocol for mapping heterogeneous wetlands with many communities from a high-resolution WV-2 satellite image. To test this developed protocol, an object-based machine learning ensemble approach was selected as a benchmark for comparison. To effectively apply the developed protocol, feature selection techniques were applied, optimal spectral and spatial features were identified, and the benefit of four additional bands of WV-2 products were evaluated. The study also applied a post classification change detection technique to delineate the change between 2017 and 2021. The developed object-based deep learning protocol has been proven superior to the object-based machine learning ensemble approach. The feed-forward neural network (FNN) deep learning classifier achieved an overall accuracy of 91.2% and 88.6% for 2017 and 2021 imagery, respectively. On the other hand, the ensemble analysis approach achieved an overall accuracy of 87.8% and 85.5% for 2017 and 2021 imagery, respectively. The FNN improved (>3%) the classification accuracy compared to the ensemble analysis, and the difference between classification results was statistically significant. The deep learning classifier not only increased the overall accuracy, but it also helped identify minor communities more accurately than the ensemble analysis technique. The additional four bands, object-based texture measures, and NDVI values of WV-2 satellite imagery showed the potential to map heterogeneous wetlands with many communities. The change map provided valuable insights into temporal changes in wetlands, which can aid in the formulation of adaptive management strategies. Exploration of automated/semi-automated deep learning methods contributed by this dissertation research will not only advance modern deep learning in wetland applications, but also assist with regional land managers to make efficient decisions by generating timely map products.

Download

Share

COinS