Incorporating multiple bands and land cover improves cross-sensor normalization

The Landsat archive and the Harmonized Landsat and Sentinel-2 (HLS) 2.0 virtual constellations enable monitoring of long-term and rapid environmental change with optical imagery, but differences in sensor design and atmospheric correction algorithms require cross-sensor normalization to ensure consistent observations. This study evaluates methods for such normalization using two global datasets: Landsat ETM+/OLI and HLS 2.0. We compared linear regression approaches and machine learning models, with and without land-cover information. For ETM+/OLI cross-sensor normalization, the XGBoost models produced the largest improvement over no normalization, reducing RMSD by 15.6%, while multiple linear regression performed best for HLS 2.0, reducing RMSD by 6.7%. Incorporating land cover further improved performance, with RMSD reductions of 20.6% for ETM+/OLI and 7.6% for HLS 2.0 compared to no cross-sensor normalization. Our results highlight the benefits of models that integrate multiple bands and landcover data, which outperformed traditional band-to-band linear regression models. This study provides practical cross-sensor normalization guidance which can benefit downstream applications such as NDVI trend analysis, land cover change detection, and generative pretrained transformer models. Building on these results, we have updated the LEOHS Python package to support XGBoost cross-sensor normalization models.