Abstract
Combining data from Sentinel-2 (MSI) and Landsat 8 (OLI) offers a powerful opportunity for high-frequency Earth observation, potentially reducing revisit times to 2–3 days. However, differences in sensor Spectral Response Functions (SRFs) introduce radiometric discrepancies. This study quantifies these differences using simulated reflectance from ASTER and Hyperion spectral libraries. Results indicate that while some bands are highly consistent, others—particularly in the visible and near-infrared (NIR) ranges—show significant variation. Notably, using Sentinel-2’s narrow NIR band (8a) instead of the broad band (8) substantially improves consistency in NDVI calculations.
Introduction
For time-sensitive applications like emergency response or agricultural monitoring, high-frequency satellite data is critical. While Sentinel-2 and Landsat 8 are both medium-resolution workhorses, their individual revisit cycles (5–10 days) are often insufficient. Harmonizing their data can provide a 2–3 day revisit cycle. However, "harmonization" is complicated by the fact that different sensors have unique "sensitivity profiles," known as Spectral Response Functions (SRFs).
Quantifying Spectral Discrepancies
The core of the challenge lies in how each sensor's filters respond to incoming light. Even when looking at the same target, MSI and OLI will report different reflectance values because their SRFs do not perfectly overlap.
As shown in the SRF comparison, there are notable shifts and width differences in the corresponding bands. The study used high-resolution "ground truth" spectra from the ASTER Spectral Library and Hyperion data archives to simulate how these SRF differences manifest as reflectance errors.
Key Findings: The NIR Advantage and Index Propagation
The research revealed that discrepancies are not uniform across all bands. Coastal Aerosol, SWIR, and Cirrus channels showed high consistency (relative difference < 5%). In contrast, the Blue, Green, Red, and broad NIR channels exhibited much higher differences.
A critical insight for vegetation monitoring is the choice of NIR band. Sentinel-2 features both a broad NIR band (Band 8) and a narrow one (Band 8a). The study found that Band 8a is a much closer match to Landsat 8’s NIR band.
As illustrated above, using Band 8a significantly reduces the discrepancy in NDVI calculations. For the ASTER dataset, this switch reduced the median relative difference in NDVI from -14.8% to -6.3%. This highlights how small band-level errors can be amplified when calculating spectral indices.
The Challenge of Universal Transformation
The study also tested linear transformation models ($OLI = Slope \times MSI + Offset$) to see if MSI data could be corrected to match OLI. A key warning for practitioners emerged: transformation models are highly sensitive to the underlying land cover. A model calibrated on one spectral library (e.g., ASTER) often performed poorly or even increased errors when applied to another (e.g., Hyperion). This suggests that there is no "universal" constant for cross-sensor calibration; corrections must likely be land-cover specific.
Conclusion
While the joint use of Sentinel-2 and Landsat 8 is promising, the differences in their spectral response functions present a significant hurdle. Achieving high consistency requires careful band selection—such as prioritizing MSI Band 8a for vegetation indices—and a recognition that simple linear transformations may not generalize across diverse landscapes. Future work must focus on developing more robust, context-aware harmonization techniques.
Summary updated