Optimizing deep neural networks for high-resolution land cover classification through data augmentation.

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Artificial Intelligence

Background:

• High-resolution land cover classification is vital for environmental monitoring.
• Manual data annotation is labor-intensive, necessitating efficient data augmentation.
• A comprehensive comparison of data augmentation for land cover classification is lacking.

Purpose of the Study:

• To systematically evaluate various data augmentation techniques for deep learning-based land cover classification.
• To identify the most effective augmentation strategies for small, high-resolution datasets.
• To provide a practical framework for optimizing land cover mapping.

Main Methods:

• Tested eight data augmentation techniques across four neural networks (U-Net, DeepLabv3+, FCN, PSPNet).
• Utilized 25 cm resolution imagery from Cantabria, Spain, creating 19 training sets and 72 models.
• Employed stratified sampling with CORINE Land Cover data for validation.

Main Results:

• Data augmentation improved model performance by up to 30%.
• The best model (DeepLabv3+ with flip, contrast, brightness) achieved 0.89 accuracy and 0.78 IoU.
• Generated land cover maps for 2014, 2017, 2019 with 87.2% accuracy.

Conclusions:

• Data augmentation is crucial for effective land cover classification with limited data.
• Specific augmentation techniques (flip, contrast, brightness) offer significant performance gains.
• The study provides a valuable resource for researchers optimizing land cover mapping workflows.