Following retrieval calibration, we will provide UQ for each stage of SIF and GPP processing from various input error sources and validation against airborne and ground data. We focus 1st on instrument (Sec 1.2.4.1) and daily integration (Sec 1.2.4.2) errors for Orbital SIF, next on input/output error sources and validation of Global SIF and Network SIF including daily integration (Sec 1.2.4.3), then on input/output error sources and validation for Upscaled GPP (Sec 1.2.4.4), and finally on the validity of SIF-GPP linearity (Sec 1.2.4.5). A such, our strategy propagates uncertainty statistics from SIF retrieval algorithm calibration to upscaled GPP.
In general, satellite precision errors have been overly optimistic, and thus will be a important focus of this proposal. OCO-2 & GOSAT SIF retrievals and 1-sigma precision errors, representing pure noise error, are well characterized and supplied with each individual sounding. For GOME-2 & SCIAMACHY, we will focus on non-fluorescing targets in semi-arid and low vegetation and cloud regions, and observe if 0-level noise corresponds with expected noise. We will use fitting residuals to estimate radiance level dependent measurement error, propagate errors linearly through the retrieval to estimate precision, then check gridbox variability due to clouds & inhomogeneities.
We will validate SIF accuracy using the RMSE to evaluate the performance and quantify the uncertainty of Network SIF against airborne and tower records. RMSE has been used recently to evaluate performance of well known ML GPP products (e.g., Tramontana et al. 2016; Jung et al. 2009). Our validation strategy leverages unique aspects of tower and airborne data: (1) Sub-grid scale acquisitions (<1 km) of airborne data to resolve SIF spatial gradients to validate spatial variability of Orbital SIF. We will aggregate airborne data into 5 km bins for each sampled biome, predict Network SIF at aggregated airborne locations, and compute biome specific RMSE of observed and predicted airborne SIF; (2) Sub-daily resolution (~1 hour) of tower data to validate daily integration and monthly sampling of Orbital SIF over multiple vegetation types. We will aggregate hourly tower data into daily averages for each tower location, predict Network SIF at tower locations and overpass times for each satellite instrument, integrate predicted SIF to daily averages as ratio of daily integrated cos(SZA) to cos(SZA) at overpass time, and compute RMSE of observed and predicted daily integrated SIF as a function of biome and month; (3) Continuous sampling (synoptic, seasonal, interannual) and bioclimatic diversity of tower data to validate annual cycles of Network SIF over multiple vegetation types and climates. Here, we will compute RMSE of Network SIF against tower observations.
Airborne data is collected with European HyPlant (Collaborator Rascher) and NASA JPL CFIS, (Co-I Frankenberg). HyPlant is designed for preparing the launch of the FLEX satellite mission. Multiple campaigns have been performed from 2012-16 sampling diverse vegetation landscapes including agricultural areas, forests, grasslands, swamps in Germany, France, Italy, Czech Republic, Finland and USA (Rascher et al. 2015). These campaigns will continue in preparation for FLEX, and synchronous flights with OCO-2 are possible. CFIS was developed by the OCO-2 team led by Co-I Frankenberg to validate OCO-2 retrievals. It combines high spectral resolution (<0.1 nm) with a wide spectral coverage (737–772nm), which is optimally suited for SIF retrievals and allows for testing general retrieval strategies at our 740nm standardized wavelength. Several flight campaigns with CFIS were carried out in 2015 & 2016 across a range of ecosystems (including crops, grassland, & forests), under-passing OCO-2 orbital tracks in the Midwest and California. Additional flights are planned in Summer 2017 in Canada and Alaska for ABoVE.
On ground level, continuous canopy SIF data will be measured from existing and and recently developed ground systems at established EC flux tower sites from KISS and FLUXNET. These instruments are developed, installed, and maintained by project PI and Co-I’s (Parazoo, Frankenberg, Magney) and collaborators (Yang, Stutz, Seibt, Keppel-Aleks, Porcar-Castell, Berry) with funding from NASA CCS and IDS projects. Together, these systems measure SIF continuously at leaf and canopy level, enabling estimates of seasonal, light and biome dependent SIF, LUE, and SIFyield needed to validate daily integration, annual cycles, and SIF-yield/LUE relationships (Sec 1.2.4.7). Two new instruments, FluoSpec2 and PhotoSpec developed by collaborators Yang, Stutz, & Seibt represent the majority of existing and planned SIF tower sites (Maine, Iowa, Michigan, Saskatchewan, Alaska, New Mexico, Michigan, Virginia). These will undergo a 2-3 week field intercomparisons at our Old Black Spruce site in Saskatchewan as part of our NASA IDS project to ensure consistency across sites and instruments. We will leverage nearby OCO-2 overpasses and CFIS field campaigns from ABoVE as part of our cal/val work.
We will also compute data fusion predictive errors for each pixel. Accumulated uncertainty statistics due to retrieval algorithm, instrument, and RMSE of observed and predicted SIF at validation sites will provide SIF UQ input to data fusion techniques.
Global SIF: Estimates of uncertainty are a byproduct of the geostatistical method. Uncertainties are nearly zero where observations are available and it increases for locations that are far away from observational locations. Measurement error is directly incorporated in the method and a non-negative solution that results in positive estimates of the gap-filled quantity can be obtained.
Network SIF: We use a Monte Carlo strategy to quantify prediction error for the RF algorithm as summarized in four main steps (Coulston et al. 2016): (1) Use bootstrap resampling to parameterize a large number of RF models. (2) For each RF (and bootstrap), create an error assessment dataset based on the datasets NOT selected for model training. (3) Utilize the error assessment dataset developed for all bootstrap replicates, perform Monte Carlo to construct a measure ? such that 95% of prediction lies within ? ∙ ???(?) (? is prediction). (4) Make a prediction for a new observation, including predictive error that represents the 95% confidence interval.
We will validate interannual variability of upscaled GPP against long term GPP tower data from FLUXNET, and seasonal and spatial variability, including biome diversity and seasonal phenology (spring onset and fall senescence) and amplitude, against upscaled flux tower data from FLUXCOM GPP.
The linear relationship between SIF and GPP observed at the ecosystem and canopy scales (Frankenberg 2011b; Guanter 2012) disappears when scaled down to leaf scale (Zhang 2016). At leaf scale, the relationship is similar to a typical leaf light response curve, whereby photosynthesis saturates at a moderate light intensity and SIF continues to scale with light intensity (Magney 2017). As such, fluorescence and CO2 yields are not constant at leaf scale (Genty 1989). CO2 yields vary with photosynthetic capacities and environmental conditions such as light, atmospheric CO2 and humidity as predicted with the Farquhar–von Caemmerer– Berry model of photosynthesis (Farquer 1980); whereas SIF yield responds to environmental conditions that affect photochemical and non-photochemical quenching (Porcar-Castell 2014). An explanation for the linearity observed at the satellite scale can be filled with mechanistic studies at smaller scales leaf, tower, and airborne levels (Zhang 2016). Leveraging data from tower studies, we will observe SIF at high temporal resolution in conjunction with changes in environmental conditions, and quantify RMSE metrics (deviations) in cases where the SIF/GPP relation becomes nonlinear under changing environmental conditions (i.e. high VPD). The SIF, GPP, ancillary data, and uncertainty metrics provided in the Upscaled GPP ESDR will provide valuable information for ecosystem modelers to parameterize non-linearities in predicted SIF/GPP relationships.