Global uncertainty assessment of vegetation indices from NASA's Harmonized Landsat and Sentinel-2 Project

NASA's Harmonized Landsat and Sentinel-2 (HLS) project recently started to produce in forward production a total of nine Vegetation Index (VI) products from the HLS version 2.0 Landsat 8–9 30 m (L30) and Sentinel-2 30 m (S30) surface reflectance data. The HLS version 2.0 dataset provides revisit observations every 1.6 days globally and every 2.2 days in the tropics (the least frequently covered latitudes), when data from four satellites (Landsat 8–9 and Sentinel-2 A/B) are available. HLS-derived VIs can provide a valuable resource for studying vegetation dynamics, including crop growth, forest loss, and disturbance severity and recovery among others. To characterize the suitability of these VIs for scientific applications, we assessed the between-sensor uncertainties for the nine HLS VI products and 12 additional ones, using VIs derived from HLS V2.0 (L30 and S30) surface reflectance for the years 2021 and 2022. A random sample of over 136 million cloud-free observations from 545 same-day L30 and S30 image pairs were selected to represent different landscapes globally in subarctic, temperate, and tropical climates. First, we evaluated between-sensor consistency for each VI derived from L30 and S30 and found high consistency (R² > 0.94) for most VIs, except for chlorophyll vegetation index (CVI, R² = 0.5). Second, we quantified the impact of potential factors on VI uncertainties using the mean absolute difference (MAD) between L30 and S30. Large view azimuth angle differences (VAD) between observation pairs (> ∼ 125°) increased MAD by ≤0.01 in most VIs. The impact on the Root Mean Square Error Interquartile Range (RMSEIQR) for these VIs varied from a decrease of 0.029 to an increase of 0.017. High solar zenith angle (SZ) (> ∼ 60°), prevalent during winter, also increased MAD by <0.07 and RMSEIQR by <0.2 for most VIs. Furthermore, one of the largest discrepancies was found in the area of terrain shadow, with a relative difference of over 20 %. The findings showed the importance of continuing HLS algorithm refinement. Finally, we analyzed VI uncertainties across VI values and for the qualitative aerosol optical depth characterization at three levels. Using VIs derived from low-level aerosols as a baseline, we assessed the impact of aerosol levels. VIs derived from moderate-level aerosol conditions closely aligned with the baseline. However, high aerosol levels introduced evident discrepancies, highlighting increased uncertainty in VIs under these conditions. Notably, even for low-level aerosol observations, uncertainties increased at VI tail values. For robust application of HLS V2.0 VIs in scientific studies, we recommend VI value ranges associated with low uncertainty. Additionally, we reported standard deviations of discrepancies, stratified by aerosol level and VI value, enabling users to account for uncertainties in their analyses, especially for VIs derived from high aerosol levels or beyond recommended ranges.