By Gerjon Schoonderwoerd
To monitor vegetation processes on the Earth Surface, Remote Sensing commonly uses Vegetation Indices (VIs). These mathematical combinations of reflectances at different wavelength bands, derived from aerial imagery, provide insight in for example vegetation health. New methods, such as Sun-Induced Fluorescence (SIF), are more directly related to actual photosynthesis and are more variable through time than traditional VIs, hence solving some of their limitations. For example, SIF responds to vegetation stress quickly, in contrast to traditional VIs. SIF involves the measurement of the emission of visible light (Fluorescence) by plants, commonly by measuring inside the narrow sub-nanometer O2-A and O2-B oxygen absorption bands, where incoming solar radiation is minimal due to atmospheric absorption. This Fluorescence is a waste product during photosynthesis, and because photosynthesis depends on sunlight, it is Sun-Induced. The amount of plant SIF is hence closely related to the amount of photosynthesis of a plant. Because the measured SIF signal is very weak (±2% of Photosynthetically Active Radiation (PAR)), currently most SIF measurements have a large footprint: point observations covering a larger area or images with a large pixel size.
As both SIF and VIs are sensitive to plant status, it was expected that there may be a correlation between the two. As different vegetation indices are sensitive to different crop properties, a mathematical combination of vegetation indices may improve predictability of SIF using VIs. The objective of this research was therefore to identify and quantify the relationship between SIF and VIs using different models.
Linear regression (1-5 parameters), linear regression with an interaction term (1-3 parameters) and k-Nearest Neighbours (kNN) regression (1-2 parameters) were applied to model SIF data using a selection of 12 broadband VIs, calculated using four bands (red, green, red edge, NIR), acquired from an Unmanned Aerial Vehicle (UAV) platform with a pixel size in the order of 5-7 cm. Model performance was quantified according to Root-Mean-Square Error (RMSE) and Coefficient of Variation (R2).
SIF observations were made using an Ocean Optics QE-Pro photospectrometer (630-800 nm at 0.14 nm FWHM (Full Width at Half Maximum)) mounted to a UAV, flown over winter wheat, potato and sugar beet fields close to Wageningen University (51.988015°N, 5.651400°E) over two days in 2018. Flight height and sensor Field of View (FOV) were used to calculate the area covered by each observation, to extract VI values over that area. A correlation between SIF and broadband VIs was found, with an R2 up to 0.7 when using linear regression on a winter wheat canopy. Model performance for potato (R2 varying between 0.403 and 0.572 on August 02) and sugar beet (0.317-0.570) is weaker. RMSE values are however around 50%, 37% and 23% of mean SIF measurement for resp. winter wheat, potato and sugar beet. In most models, there is an underestimation of high SIF values. kNN regression and interaction models generally perform similar or worse than linear regression. Linear regression is considered most suitable for predicting SIF. Although some VIs perform stronger than others in some cases, no consistent best VI has been found for the crops considered.
It was found that several error sources may have weakened accuracy. These consist of e.g. GPS accuracy (0.65 m for SIF and ≥0.26 m for VI imagery) and sensor movement compared to a mean FOV of 2.22 m and measurement accuracy, estimated at a range of 0.16 mW/m2/nm/sr, for measurements ≤1.5 m apart.
With R2 values up to 0.7, there is a relationship between SIF and VIs. Questions such as: “To what extent are SIF model residuals caused by differences between VIs and SIF signals, and to what extent by measurement errors?” still need to be answered before drawing conclusions about how strong this relationship really is. Further research is hence needed.