Generative Framework Approach to Match Landsat and Sentinel-2 Data

Linear regression and histogram matching-based techniques have been widely used to minimize the surface reflectance difference between two similar satellite observations such as Landsat-8/9 and Sentinel-2A/B products. However, regionally or globally derived conversion factors may not be suitable for all land cover types and locations, resulting in noticeable residual differences between the sensors. Generative Adversarial Network (GAN) has shown promise in the field of image processing for domain or style transfer. In this work we aim to minimize the surface reflectance difference between Landsat and Sentinel-2 products based on GAN.This work selected 26 pairs of same-day Landsat and Sentinel-2 30-meter spatial resolution surface reflectance images in the green spectral band from NASA’s Harmonized Landsat/Sentinel-2 project (HLS), with each pair having over 90% spatial overlap. Upstream pre-processing steps included atmospheric correction, cloud masking and BRDF normalisation. The generator architecture was based on U-Net and discriminator as PatchGAN. GAN was trained for 43000 iterations. Finally, the Landsat images generated from the Sentinel-2 images by GAN are compared to the original Landsat 8/9 images in terms of SSIM and MSE metrics. Result showed a higher SSIM score for the generated data, which can be interpreted to mean that the generated images were closer to the real Landsat images. And, overall MSE for generated data was lower than that for the original Sentinel-2 images.This study, for the first time, reports a GAN-based spatial matching between Landsat8/9 and Sentinel-2 surface reflectance images. The results indicate that this approach has the potential to map data between the two satellite images with reasonable accuracy. This method may prove to be more robust and could be applied globally, potentially replacing the previous approach of simulated data. The method can be further extended with more data and in other spectral bands and potentially be used in HLS processing.

I. Introduction

Remote sensing data is commonly used to build time series to study changes in regions of the Earth. Hundreds of satellites are currently in orbit gathering data at different times and wavelengths. Remote sensing data users have historically taken on the difficult task of normalizing datasets with similar properties to create denser time series when needed. More recently, data providers have started highly accurate multisensor harmonization efforts to reduce that necessary effort for users and create larger datasets for deep learning-based earth science applications. This research aims to demonstrate the value provided by machine learning techniques for multisensor harmonization.