A global spatially Continuous Solar Induced Fluorescence (CSIF) dataset using neural networks

Abstract

Abstract. Satellite-retrieved Solar Induced Chlorophyll Fluorescence (SIF) has shown great potential to monitor the photosynthetic activity of terrestrial ecosystems. However, several issues, including low spatial and temporal resolution of 10 the gridded datasets and high uncertainty of the individual retrievals, limit the applications of SIF. In addition, inconsistency in measurements footprints also hinder the direct comparison between gross primary production (GPP) from eddy covariance (EC) flux towers and satellite-retrieved SIF. In this study, by training a neural network (NN) with surface reflectance from the MODerate-resolution Imaging Spectroradiometer (MODIS) and SIF from Orbiting Carbon Observatory-2 (OCO-2), we generated two global spatially continuous SIF (CSIF) datasets at moderate spatio-temporal resolutions (0.05 degree 4-day) 15 during 2001-2016, one for clear-sky conditions and the other one in all-sky conditions. The clear-sky instantaneous CSIF (CSIFclear-inst) shows high accuracy against the clear-sky OCO-2 SIF and little bias across biome types. The all-sky daily average CSIF (CSIFall-daily) dataset exhibits strong spatial, seasonal and interannual dynamics that are consistent with daily SIF from OCO-2 and the Global Ozone Monitoring Experiment-2 (GOME-2). An increasing trend (0.39%) of annual average CSIFall-daily is also found, confirming the greening of Earth in most regions. Since the difference between satellite 20 observed SIF and CSIF is mostly caused by the environmental down-regulation on SIFyield, the ratio between OCO-2 SIF and CSIFclear-inst can be an effective indicator of drought stress that is more sensitive than normalized difference vegetation index and enhanced vegetation index. By comparing CSIFall-daily with gross primary production (GPP) estimates from 40 EC flux towers across the globe, we find a large cross-site variation (c.v. = 0.36) of GPP-SIF relationship with the highest regression slopes for evergreen needleleaf forest. However, the cross-biome variation is relatively limited (c.v. = 0.15). These two 25 continuous SIF datasets and the derived GPP-SIF relationship enable a better understanding of the spatial and temporal variations of the GPP across biomes and climate.

NIRvP: a robust structural proxy for sun-induced chlorophyll fluorescence and photosynthesis across scales

“…Therefore, detailed analyses at finer spatial scales are needed. Such studies should consider using enhanced SIF products with higher signal quality and spatio-temporal resolution 43,47 or higher quality SIF retrievals that will be available in the future from new satellites 2,48 .…”

“…S7). Both for TROPOMI SIF retrievals and the machine learning product CSIF 43 , NIRVP had the highest spatial and temporal correlations to SIF followed by EVI2P, FCVIP, APAR and NDVIP (Fig. S7).…”

“…As additional test, we compared MODIS NIRVP with CSIF 43 , which is a machine learning product based on OCO-2 SIF retrievals, MODIS reflectance and fPAR products. Since the original version of the product did not cover 2018, we used the recently updated version 2 of the product.…”

Canopy structure explains the relationship between photosynthesis and sun-induced chlorophyll fluorescence in crops

“…It seems, therefore, that for APAR × f esc -based GPP estimation, an ecosystem-dependent slope has to be applied as was partly done in Badgley et al (2019). This is particularly relevant for evergreen needleleaf forests (ENF) that have a much lower NIR reflectance despite rather high GPP during the growing period, as can be inferred also from previous studies investigating SIF obs -GPP relationships (Sun et al, 2018;Yao Zhang et al, 2018a).…”

Upscaling Solar-Induced Chlorophyll Fluorescence from an Instantaneous to Daily Scale Gives an Improved Estimation of the Gross Primary Productivity

“…Considering the variations in FPAR and fluorescence efficiency with the vegetation growing in one month, SIF products at higher temporal resolutions (e.g., daily or weekly) can be also derived using the data-driven method by integrating reflectance-based products (reflectance, vegetation index, FPAR) and PAR data. Similar to the approach in the studies of References [58,59], a neural network can first be trained using the instantaneous SIF observations and the instantaneous driven-variables (including PAR, reflectance-based products); then the high-temporal resolution SIF at a daily or weekly resolution can be calculated by using the trained neural network and its key driven variables, including reflectance-based vegetation products (reflectance, vegetation index, FPAR) and the daily or weekly PAR dataset.…”