Optimizing Satellite Imagery Datasets for Enhanced Land/Water Segmentation.

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Machine Learning

Background:

• Accurate land/water segmentation is crucial for environmental monitoring.
• Deep learning models require high-quality annotated datasets for optimal performance.
• Manual dataset optimization is time-consuming and prone to errors.

Purpose of the Study:

• To develop an automated procedure for optimizing land/water segmentation datasets.
• To enhance the quality of multispectral satellite image annotations.
• To improve the performance of deep learning models in remote sensing tasks.

Main Methods:

• Utilized the Normalized Difference Water Index (NDWI) with a variable threshold.
• Developed an automated system to assess and exclude low-quality image annotations.
• Applied the method to optimize publicly available SWED and SNOWED datasets.

Main Results:

• Deep learning models trained on optimized datasets showed superior performance compared to baseline datasets.
• Achieved up to a 10% increase in mean intersection over union (mIoU).
• Demonstrated improved segmentation accuracy despite a reduction in dataset size.

Conclusions:

• The automated NDWI-based methodology offers a scalable solution for dataset optimization.
• Enhancing dataset quality significantly improves deep learning model performance in remote sensing.
• The approach is promising for environmental monitoring and other geospatial applications.