The successful operation of airborne and space-based spectrometers in recent years holds the promise to map solar-induced fluorescence (SIF) accurately across the globe. Machine learning (ML) can play an important role in this effort, but its application to SIF retrieval methods is in part hindered by the need for time-consuming radiative transfer modelling to account for atmospheric effects. In this work, we address this difficulty and develop a fast and accurate physics-based ML emulator of at-sensor radiances around the O $_{2}$ -A absorption band for the space-based DESIS and the airborne HyPlant spectrometers. Different ML models are trained on an extensive set of simulated spectra encompassing a wide range of atmosphere, geometry, surface and sensor configurations. A fourth-degree polynomial model is found to perform best, presenting errors at or below 10% of typical SIF at-sensor radiances and a prediction time per sample spectrum of 10-20 $\mu$ s. Using data acquired with the HyPlant instrument, the proposed model is also shown to be able to match very closely the measured spectra. We illustrate how to improve further the accuracy of the emulator and how to generalize it to other sensors using the particular case of ESA's FLEX space mission. Our findings suggest that physics-based emulators can be efficiently used for the development of ML-based SIF retrieval methods by generating large training data sets in short time and by enabling a fast simulation step for self-supervised retrieval schemes.
Accurate estimation of solar-induced fluorescence (SIF) from passively sensed hyperspectral remote sensing data has been identified as instrumental in assessing the photosynthetic activity of plants for various scientific and ecological applications at various spatial scales. Different techniques to derive SIF have been developed over the last decades. In view of ESA's upcoming Earth Explorer satellite mission FLEX aiming to provide high-quality global imagery for SIF retrieval an increased interest is placed in physical approaches. We present a novel method to retrieve SIF in the O 2 -A absorption band of hyperspectral imagery acquired by the HyPlant sensor system. It aims at a tight integration of physical radiative transfer principles and self-supervised neural network training. To this end, a set of spatial and spectral constraints and a specific loss formulation are adopted. In a validation study we find good agreement between our approach and established retrieval methods as well as with in-situ top-of-canopy SIF measurements. In two application studies, we additionally find evidence that the estimated SIF (i) satisfies a first-order model of diurnal SIF variation and (ii) locally adapts the estimated optical depth in topographically variable terrain.
We provide the first method allowing to retrieve spaceborne SIF maps at 30 m ground resolution with a strong correlation ( $$r^2=0.6$$ ) to high-quality airborne estimates of sun-induced fluorescence (SIF)ṠIF estimates can provide explanatory information for many tasks related to agricultural management and physiological studies. While SIF products from airborne platforms are accurate and spatially well resolved, the data acquisition of such products remains science-oriented and limited to temporally constrained campaigns. Spaceborne SIF products on the other hand are available globally with often sufficient revisit times. However, the spatial resolution of spaceborne SIF products is too small for agricultural applications. In view of ESA’s upcoming FLEX mission we develop a method for SIF retrieval in the O $$_2$$ -A band of hyperspectral DESIS imagery to provide first insights for spaceborne SIF retrieval at high spatial resolution. To this end, we train a simulation-based self-supervised network with a novel perturbation based regularizer and test performance improvements under additional supervised regularization of atmospheric variable prediction. In a validation study with corresponding HyPlant derived SIF estimates at 740 nm we find that our model reaches a mean absolute difference of $$0.78 \, \, \mathrm {mW\, nm^{-1} \, sr^{-1} \, m^{-2}}$$ .
The retrieval of sun-induced fluorescence (SIF) from hyperspectral imagery requires accurate atmospheric compensation to correctly disentangle its small contribution to the at-sensor radiance from other confounding factors. In spectral fitting SIF retrieval approaches this compensation is estimated in a joint optimization of free variables when fitting the measured at-sensor signal. Due to the computational complexity of Radiative Transfer Models (RTMs) that satisfy the level of precision required for accurate SIF retrieval, fully joint estimations are practically inachievable with exact physical simulation. We present in this contribution an emulator-based spectral fitting method neural network (EmSFMNN) approach integrating RTM emulation and self-supervised training for computationally efficient and accurate SIF retrieval in the O 2 -A absorption band of HyPlant imagery. In a validation study with in-situ top-of-canopy SIF measurements we find improved performance over traditional retrieval methods. Furthermore, we show that the model predicts plausible SIF emission in topographically variable terrain without scene-specific adaptations. Since EmSFMNN can be adapted to hyperspectral imaging sensors in a straightforward fashion, it may prove an interesting SIF retrieval method for other sensors on airborne and spaceborne platforms.