The emission of solar-induced chlorophyll fluorescence (F) is a pivotal process to infer vegetation health and functioning that can be monitored by remote sensing. However, most of the current remote sensing methods retrieve only F at top-of-canopy level, therefore making the link with physiological processes occurring at photosystem level not trivial. In this study, we develop a novel machine learning Fourier (phasor)-based algorithm to retrieve F both at canopy level and after considering the reabsorption (i.e. photosystem level), consistently with relevant biophysical variables, exploiting the canopy apparent reflectance spectra (R_app). In particular, R_app is divided in consecutive spectral windows, where the Discrete Fourier Transform (DFT) is computed. Then, the DFT results in each window are exploited to estimate the fluorescence spectra and biophysical parameters, together with their uncertainties, by means of a supervised machine learning algorithm coupled to a statistical-based retrieval pipeline. The algorithm has been trained through synthetic R_app spectra, obtained from simulations based on a Radiative Transfer (RT) model. As a proof of concept, the theoretical approach is then applied to experimental data, acquired both from crops and forests, at close and high soil-sensor distance respectively, to evaluate the retrieval accuracy of biophysical and F parameters. In particular, for the first time R_app is used to extract the temporal evolution of F at canopy and photosystem levels and its quantum efficiency together with different biophysical variables, during the growing season of two agricultural crops. Furthermore, tower-based solar-induced fluorescence measurements in a deciduous forest are exploited to evaluate the performance of our algorithm when the atmospheric reabsorption and scattering are not negligible. The reliability of the proposed method is evaluated through a comparison with F spectra extracted from the state of the art SpecFit retrieval algorithm. This work promises a substantial advance toward a new accurate retrieval method for fluorescence signals and biophysical parameters at canopy and photosystem levels.

太阳诱导的叶绿素荧光 (F) 的发射是推断植被健康和功能的关键过程，可以通过遥感进行监测。然而，当前大多数遥感方法仅在冠层顶部检索 F，因此与在光系统水平发生的生理过程的联系并非微不足道。在这项研究中，我们开发了一种新的基于机器学习傅立叶（相量）的算法，以在冠层水平和考虑再吸收（即光系统水平）后检索 F，与相关的生物物理变量一致，利用冠层表观反射光谱（R _app）。特别是，R_应用程序被划分为连续的光谱窗口，其中计算离散傅里叶变换 (DFT)。然后，通过与基于统计的检索管道耦合的监督机器学习算法，利用每个窗口中的 DFT 结果来估计荧光光谱和生物物理参数及其不确定性。该算法已通过合成R _app光谱进行了训练，该光谱是从基于辐射转移 (RT) 模型的模拟中获得的。作为概念验证，该理论方法随后应用于分别在近距离和高土壤传感器距离处从农作物和森林获取的实验数据，以评估生物物理和 F 参数的检索准确性。特别是第一次使用R_应用程序用于在两种农作物的生长季节提取 F 在冠层和光系统水平的时间演化及其量子效率以及不同的生物物理变量。此外，当大气再吸收和散射不可忽略时，利用落叶林中基于塔的太阳诱导荧光测量来评估我们的算法的性能。通过与从最先进的 SpecFit 检索算法中提取的 F 谱进行比较来评估所提出方法的可靠性。这项工作有望在树冠和光系统水平上对荧光信号和生物物理参数的新的准确检索方法取得重大进展。