Deep Learning-Based Stripe Restoration Algorithm for Landsat 7

Abstract

Since the failure of the Scan Line Corrector (SLC) on the Landsat 7 ETM+ sensor in 2003, approximately 22% of pixels in acquired images are missing in a regular stripe pattern, severely limiting their application in time-series analysis and land cover monitoring. To address the limitations of traditional methods in modeling complex land surfaces and their tendency to introduce spectral distortions, this paper proposes a stripe-aware deep learning restoration network called SINet (Stripe Inpainting Network). The network leverages a Stripe Attention Module (SAM) to exploit the geometric prior of stripes and aggregate features along the stripe direction, and a Spectral Reconstruction Module (SRM) to model multi-band correlations and preserve spectral fidelity. A hybrid loss function combining pixel loss, perceptual loss, spectral angle loss, and gradient loss is designed. Experiments conducted in the Zhengzhou-Kaifeng-Xuchang junction area of Henan Province show that SINet achieves a PSNR of 38.2 dB, an SSIM of 0.978, and an SAM of 1.82° on simulated data, significantly outperforming baseline methods such as nearest neighbor interpolation, bilinear interpolation, and local histogram matching. When applied to land cover classification, the overall accuracy improves from 74.3% (uncorrected) to 88.5%, approaching the 90.2% accuracy of original Landsat 8 imagery. This study provides an effective solution for restoring historical Landsat 7 data.