Evaluation of spatial and temporal variability in Sentinel-2 surface reflectance on a rice paddy landscape

Abstract

• Pixelwise evaluation of Sentinel-2 reflectance against drone and tower observations. • Point spread function constrained the spatial representation of Seninel-2 imagery. • Image contrast reduction increased the RMSE of Sentinel-2 reflectance by up to 7.6 %. • Footprint mismatching contributed to an error source of 2.1–11.8 % in RMSE. High spatial resolution spaceborne remote sensing systems provide a new data source for agricultural applications. As a key deliverable, surface reflectance (SR) enables immediate and non-destructive estimation of crop status, thus the demand for reliable pixelwise SR is increasing. However, the evaluations are typically conducted on pseudo-invariant areas and the reliability of pixelwise SR has not been thoroughly examined over heterogenous, dynamic surfaces. In this study, we evaluated pixelwise Sentinel-2 (S2) SR on a rice paddy landscape across seasons using drone-based hyperspectral images and tower-based continuous hyperspectral observations as the ground references. We also examined the impact of spatial and atmospheric properties on S2 SR. Overall, S2 SR showed strong linear relationships with the ground references (the overall R 2 > 0.8). The residual errors were influenced by sub-pixel geolocation errors (0.01–0.017 (2.1–11.8 %)), a widen PSF (0.007 (7.6 %) for red) and underestimated AOT retrievals (0.027 (40.7 %) for blue). Notably, atmospheric adjacency effects broadened the PSF, causing the consistency of S2 with the ground reference image to depend on the landscape's heterogeneity. Our findings outlined the key factors contributing to uncertainties in S2 SR, which could affect downstream products like vegetation indices and leaf area index. Considering these factors would enhance remote sensing of landscapes with high contrast in reflectance and elevated aerosol loadings, such as urban, savanna, wetland and dry agricultural land.