A GIS-BASED EXPERT SYSTEM TO IMPROVE THE ACCURACY OF WETLAND CLASSIFICATION
Analytics
95 views ◎53 downloads ⇓
Abstract
Wetlands play a critical role in our natural environment, such as improving water quality, controlling erosion and flooding, and protecting biodiversity. To better protect wetland systems, a comprehensive knowledge of their spatial distribution is important to minimize potentially devastating impacts and help with improving wetland function. For instance, accurate wetland delineation guides the mitigation plan in the transportation and construction work to protect wetlands as the US 1970 National Environmental Policy Act (NEPA) requires. Determining the precise location and extent of wetlands across large-scale regions requires a substantial amount of fieldwork. Therefore, automatic image classification with the support of remote sensing data has become a trend in studying wetland distribution. Current wetland classification studies leverage statistical or machine learning methods to build spatial models based upon the training dataset. They apply these models to predict the occurrence of wetlands, which can later be evaluated through actual fieldwork. However, current studies often face challenges introduced by the data quality. For example, the process of collecting data may introduce inaccuracy and the samples may not reflect the characteristics of the objective region. These factors have the potential to bias the identification of boundaries among different wetland types. Therefore, a flexible framework that can take into consideration the data quality and produce promising results for different scenarios is necessary. This dissertation focuses on the development of such a wetland type classification framework that can predict the spatial distribution of different types of wetlands in North Carolina. The overall objective is to build a robust and reliable expert system that can accurately classify wetland types, using training datasets of various quality. To be more specific, this system should be able to tolerate unbalanced and less representative data samples in the training data. To improve the quality of the classification model, I use various data sources to generate detailed topographic information, such as high-resolution Light Detection and Ranging (LiDAR) data, satellite data, and soil data. I also develop an integrated method to combine advantages of different models and compensate for unbalanced and limited data samples. Lastly, I construct a GIS-based orchestration system to facilitate the replication of the modelling process in a different region. Leveraging this framework, I conduct different experiments to test the model performance responding to various sampling conditions. The results reveal that the machine learning based methods mainly rely on the quality of the data over the quantity. Under a representative distribution, a sampling data set using five percent of the population proves as accurate as a sampling data set using eighty percent. In the opposite scenario, the proposed integrated method can produce better prediction accuracy than any individual model.