Guiding Labelling Effort for Efficient Learning With Georeferenced Images

Area of Science:

• Computer Science
• Geospatial Analysis
• Machine Learning

Background:

• Training deep learning models like convolutional neural networks (CNNs) for georeferenced imagery often requires extensive labeled data.
• Limited transferability of learning across different georeferenced datasets necessitates dataset-specific training.
• Existing methods like transfer learning and active learning can still be annotation-intensive.

Purpose of the Study:

• To develop a novel semi-supervised learning method to reduce labeling effort for training CNNs on georeferenced imagery.
• To enable efficient, per-dataset training of CNNs in domains with poor cross-dataset transferability.
• To demonstrate the method's effectiveness across diverse georeferenced image datasets.

Main Methods:

• A semi-supervised learning approach utilizing a location-guided autoencoder to identify representative image subsets from unlabeled data.
• Latent space representation is used to guide the selection of informative samples for training.
• The method was evaluated on four distinct ground-truthed datasets of georeferenced environmental monitoring images (aerial and seafloor).

Main Results:

• Significant efficiency gains in labeling effort were observed across all tested aerial and seafloor image datasets.
• The method achieved equivalent accuracy to conventional training with an order of magnitude fewer annotations.
• With only 40 prioritized annotations, the method reached 85% of the accuracy achieved with approximately 10,000 human annotations.

Conclusions:

• The novel semi-supervised learning method substantially reduces the annotation burden for training CNNs on georeferenced imagery.
• The approach proves beneficial across various application domains, particularly for datasets with imbalanced class distributions or rare classes.
• This method offers a practical solution for efficient deep learning model training in specialized georeferenced data analysis.