Comparing Pixel- and Object-Based Approaches in Effectively Classifying Wetland-Dominated Landscapes.

Area of Science:

• Ecology
• Remote Sensing
• Geospatial Analysis

Background:

• Wetland ecosystems provide vital ecological services but face significant global losses.
• Effective wetland management and monitoring rely on accurate satellite remote sensing and classification.
• Selecting optimal classification approaches (pixel-based vs. object-based) is crucial for wetland habitat description.

Purpose of the Study:

• To compare pixel-based and object-based image analysis (OBIA) methods for wetland classification.
• To evaluate the performance of parametric (ISODATA, ML) and non-parametric (RF) algorithms in wetland mapping.
• To determine the most suitable approach for classifying wetland-dominated landscapes.

Main Methods:

• Conducted pixel-based and OBIA using ISODATA, ML, and RF algorithms on Quickbird satellite imagery.
• Analyzed multispectral bands and spatial/spectral metrics, including texture and vegetation indices.
• Utilized field-based data for training and validation, comparing three- and five-layer predictor stacks.

Main Results:

• Pixel-based random forest (RF) with a three-layer stack achieved the highest initial accuracy (87.9%).
• Object-based RF (OBIA) accuracy increased with five layers (90.4%), but required significant user input and resources.
• No statistically significant difference in overall accuracy was found among pixel-based ML, RF, and OBIA RF classifiers.

Conclusions:

• Pixel-based RF approaches are generally satisfactory and efficient for classifying wetland-dominated landscapes.
• OBIA, while potentially useful, demands substantial resources and user-defined parameters like segmentation scale.
• The study highlights the effectiveness of parsimonious predictor sets in achieving high wetland classification accuracy.