A Single Image Deep Learning Approach to Restoration of Corrupted Landsat-7 Satellite Images

Area of Science:

• Earth and Space Sciences
• Computer Science
• Agricultural Science

Background:

• Remote sensing is vital for agriculture, with Landsat-7 providing crucial multi-spectral imagery.
• The Landsat-7 scan line corrector (SLC) malfunction since May 2003 results in significant data gaps (up to 22% per scene).
• Existing gap-filling methods struggle with the extent of data loss.

Purpose of the Study:

• To introduce and evaluate a novel single-image gap-filling method for corrupted remote sensing data.
• To leverage the deep image prior (DIP) technique for reconstructing missing data in Landsat-7 scenes.
• To quantitatively compare the DIP approach against classical gap-filling methods.

Main Methods:

• A single-image restoration approach utilizing the deep image prior (DIP) method.
• Testing DIP's efficacy on remote sensing scenes with varying degrees of data corruption.
• Comparative analysis with established single-image gap-filling techniques.

Main Results:

• The DIP approach demonstrated superior quantitative performance, achieving an R² of 0.812, compared to 0.685 for classical methods.
• The method shows robustness, with performance influenced by the number of corrupted pixels.
• Restoration quality remains high even with substantial data loss.

Conclusions:

• The deep image prior method offers a significant advancement in filling data gaps for corrupted remote sensing imagery.
• This technique provides a more effective solution than traditional methods for Landsat-7 data.
• The approach has the potential to enhance various agricultural studies and applications reliant on remote sensing data.