A 1.2 Billion Pixel Human-Labeled Dataset for Data-Driven Classification of Coastal Environments

Area of Science:

• Earth and Environmental Sciences
• Remote Sensing
• Geospatial Analysis

Background:

• Coastlines are complex, dynamic systems influenced by natural and human factors.
• Accurate mapping of coastal environments requires frequent observations from remote sensing.
• Machine learning models for image segmentation need extensive labeled datasets.

Purpose of the Study:

• To introduce "Coast Train," a novel, large-scale, multi-labeler dataset for coastal environment image segmentation.
• To facilitate the development and validation of machine learning models for coastal dynamics research.
• To enable precise spatio-temporal mapping of coastal landforms, ecosystems, and human interventions.

Main Methods:

• Development of the "Coast Train" dataset comprising orthomosaic and satellite imagery of diverse coastal regions.
• Implementation of a human-in-the-loop tool for efficient and reproducible image segmentation and labeling.
• Collection of 1.2 billion labeled pixels across over 3.6 million hectares, with multiple labelers for agreement quantification.

Main Results:

• The "Coast Train" dataset provides extensive, diverse, and accurately labeled imagery for machine learning applications.
• The human-in-the-loop tool ensures rapid and reproducible image segmentation, crucial for large datasets.
• Quantifiable pixel-level agreement among multiple labelers enhances dataset reliability.

Conclusions:

• "Coast Train" significantly advances the capacity for machine learning-based coastal zone monitoring and analysis.
• The dataset and associated tools support improved understanding and prediction of coastline changes.
• This resource is vital for researchers studying coupled human-natural systems in coastal environments.