Land Use Land Cover Labeling of GLOBE Images Using a Deep Learning Fusion Model.

Area of Science:

• Earth Science
• Computer Science
• Remote Sensing

Background:

• Traditional land use land cover classification relies on satellite imagery.
• High-resolution aerial imagery offers greater detail but presents labeling challenges.
• Citizen science initiatives like the Global Learning and Observations to Benefit the Environment (GLOBE) program collect diverse, real-world land cover images.

Purpose of the Study:

• To address the challenge of accurate land cover labeling for citizen science image datasets.
• To develop an automated method for classifying land cover from cluttered, multi-view RGB images.
• To improve the reliability and accuracy of land cover databases derived from diverse sources.

Main Methods:

• An integrated deep learning architecture combining Unet and DeepLabV3 for initial image segmentation.
• A weighted fusion model to combine segmentation labels from multiple images of the same site.
• Training deep learning models using labeled land cover images from the GLOBE database.
• Utilizing five directional views (north, south, east, west, down) for robust label assignment.

Main Results:

• Successfully labeled 2916 GLOBE land cover images with high accuracy.
• Achieved 90.97% label accuracy using the proposed integrated model with minimal human intervention.
• Demonstrated the effectiveness of the fusion model in handling image clutter and pixel uncertainties.

Conclusions:

• The developed integrated architecture significantly enhances land cover image labeling accuracy.
• The weighted fusion model provides a robust solution for classifying large RGB image databases.
• This approach facilitates the creation of more accurate and extensive land cover datasets from citizen science contributions.