A more precise retrieval of sun-induced chlorophyll fluorescence from satellite data using artificial neural networks

In recent years, sun-induced chlorophyll fluorescence (SIF) retrieved from satellite platforms has been demonstrated to be a good proxy of gross primary production (GPP). However, existing data-driven methods based on singular value decomposition (SVD) commonly lead to high retrieval noise for single SIF observations, given the fact that SIF often has a low signal-to-noise ratio. Spatial and/or temporal aggregation has typically been used to reduce these noises. Such aggregation may diminish the effective spatial and/or temporal resolutions of current SIF products but potentially exacerbate uncertainty in interpreting SIF data. To address this issue, this study proposes a more precise data-driven method for retrieving SIF using an artificial neural network (ANN), which was trained using spatially aggregated SIF as the response variable and radiance in the spectral region of 743–758 nm as the explanatory variable. The feasibility of the ANN-based SIF retrieval method was first demonstrated using model simulations based on SCOPE and MODTRAN. Then, the ANN models were trained using OCO-2/3 SIF and TROPOMI radiance after careful matching of the overpass time and “sun-target-viewing” geometry. OCO-2/3 SIF, retrieved using high-spectral-resolution data within a narrow spectral window, is considered accurate and OCO-2 SIF was further validated by airborne SIF measurements. The resulting ANN model led to a high retrieval accuracy for SIF with an R2 of 0.85 and an RMSE of 0.217 mW∙m−2∙nm−1∙sr−1. Finally, the ANN-based method was adopted to produce the global TROPOMI SIF from May 2018 to December 2024. Assuming that the RMSE is representative of the average retrieval noises of single ANN-based SIF, the retrieval noises of ANN-based TROPOMI SIF were approximately half of the reported noises of SVD-based TROPOMI SIF. The advantage of the low retrieval noise of ANN-based SIF was proven in interpreting the seasonal patterns of SIF and estimating GPP compared with SVD-based SIF. This study provides a new insight into SIF retrieval, and the resulting ANN-based SIF product would contribute to better global carbon cycle observations.