Classification of Compressed Remote Sensing Multispectral Images via Convolutional Neural Networks

Area of Science:

• Earth observation
• Remote sensing
• Data compression

Background:

• Multispectral sensors generate high-dimensional Earth observation data, requiring compression for storage and transmission.
• Lossy compression introduces artifacts that can degrade the utility of remote sensing data, particularly for downstream machine learning tasks.
• Existing compression methods struggle to preserve the integrity of complex, multi-dimensional data structures.

Purpose of the Study:

• To develop a resource-efficient compression scheme for multispectral Earth observation data.
• To investigate the impact of data compression on deep learning-based land-cover classification.
• To propose a novel method for mitigating compression-induced artifacts and handling missing data.

Main Methods:

• Encoding multispectral observations into high-order tensor structures.
• Applying quantized low-rank tensor completion for data compression and recovery.
• Evaluating compression performance using image quality metrics and land-cover classification accuracy.
• Utilizing ESA Sentinel-2 satellite data for experimental analysis.

Main Results:

• Minimal compression significantly degrades the performance of deep learning land-cover classification models.
• The proposed quantized low-rank tensor completion scheme effectively reduces compression artifacts.
• The recovery scheme demonstrates robustness in addressing missing observations caused by cloud cover.
• The method preserves classification accuracy even under significant compression ratios.

Conclusions:

• Data compression is essential for Earth observation but poses challenges for machine learning applications.
• Quantized low-rank tensor completion offers an effective solution for compressing multispectral data while preserving its analytical value.
• The proposed method enhances the reliability of remote sensing data for critical applications like land-cover mapping.