GIScience & Remote Sensing (Dec 2023)

Evaluating the contribution of Sentinel-2 view and illumination geometry to the accuracy of retrieving essential crop parameters

  • Mahlatse Kganyago,
  • Georgios Ovakoglou,
  • Paidamwoyo Mhangara,
  • Clement Adjorlolo,
  • Thomas Alexandridis,
  • Giovanni Laneve,
  • Juan Suarez Beltran

DOI
https://doi.org/10.1080/15481603.2022.2163046
Journal volume & issue
Vol. 60, no. 1

Abstract

Read online

Wide field-of-view (FOV) sensors such as Sentinel-2 exhibit per-pixel view and illumination geometry variation that may influence the retrieval accuracy of essential crop biophysical and biochemical variables (BVs) for precision agriculture. However, this aspect is rarely studied in the existing literature. Hence, the current study aimed to evaluate the contribution of view and illumination geometries to the accuracy of retrieving Leaf Chlorophyll a and b (LCab), Canopy Chlorophyll Content (CCC), and Leaf Area Index (LAI) using the Random Forest (RF). The experiments were performed on various input variable scenarios where per-pixel geometric covariates, i.e. View and Sun Zenith Angles (VZA and SZA, respectively), and Relative Azimuth Angle (RAA), are excluded and included in spectral bands (SB) and spectral vegetation indices (SVIs), respectively, in two semi-arid areas. The results showed that spectral bands or vegetation indices combined with geometric covariates improved the R2 by 10–15% for LAI and 3–5% for CCC. In contrast, negligible improvements of 1–2% were achieved for LCab with cross-validation test data and independent held-out dataset, respectively. In agreement with previous studies, VZA and SZA were among the topmost influential variables in the RF models for estimating LAI, LCab, and CCC. Collectively, per-pixel geometric variables explained more than 30% of the variability in surface reflectance for all Sentinel-2 spectral bands (p < 2.2e-16). Overall, the results showed that incorporating geometric covariates improved the accuracy of retrieving BVs; thus, it provided additional information that improves the predictive power of SB and SVIs. The significant benefits of the geometric variables were mainly realized for canopy-level BVs (i.e. LAI and CCC) than for LCab. Therefore, it is recommended to incorporate per-pixel view and illumination geometry in estimating LAI and CCC, especially when using wide-view sensors such as Sentinel-2. However, further testing in different phenology, site and acquisition conditions is needed to confirm the contribution of the geometric covariates to facilitate reliable retrieval of BVs from remotely sensed data and aid better agronomic decisions.

Keywords