Extraction of sub-pixel C3/C4 emissions of solar-induced chlorophyll fluorescence (SIF) using artificial neural network

Solar-induced chlorophyll fluorescence (SIF) is a signal directly and functionally related to photosynthetic activity and thus holds great promise for large-scale agricultural monitoring. However, the coarse spatial resolution of existing satellite SIF observations usually consist of mixed SIF signals contributed by different crop types with distinct phenology (modulated by management practices) and varying SIF emission capacities, which impedes effective utilization of existing SIF records for large-scale agricultural applications. This study makes the first effort to overcome this challenge by developing a sub-pixel SIF extraction framework for corn and soybean in the US Corn Belt as a case study. Here we developed a machine learning (ML) based sub-pixel SIF extraction framework using Orbiting Carbon Observatory 2 (OCO-2), whose high-resolution SIF acquired along orbits at nadir enables the identification of relatively pure pixels dominated by single corn or soybean crops, facilitating validation of the developed framework. To achieve this, we first generated artificially mixed SIF pixels from pure pixels randomly weighted by fractional area coverage. We then employed a standard feed forward artificial neural network (ANN) to estimate sub-pixel SIF for corn and soybean respectively, using the following predictors: total mixed SIF, spectral reflectance of corn/soybean (from Moderate Resolution Imaging Spectroradiometer MODIS), and the fractional area coverage of corn/soybean (derived from CropScape-Cropland Data Layer). Our results demonstrated that the estimated sub-pixel SIF could successfully reproduce the original pure SIF values constituting the mixed pixel, with a normalized root mean squared error (NRMSE) of <10% during the peak growing season. We further demonstrated that this ANN-based framework substantially outperforms the parsimonious linear extraction methods. This developed sub-pixel SIF extraction framework was then applied to generate regional-scale SIF maps for corn and soybean at 0.05° in the US Midwest. Although tested for corn and soybean only, the developed framework has the potential to resolve sub-pixel SIF of more endmembers from coarse SIF observations.