Remote sensing employs solar-induced chlorophyll fluorescence (SIF) as a proxy for photosynthesis from field to airborne and satellite sensors. The investigation of SIF offers a unique way of studying vegetation functioning from the local to the global scale. However, the passive, optical retrieval of the SIF signal is still challenging. Common retrieval approaches extract the SIF infilling directly from atmospheric oxygen bands in down-welling and up-welling radiance. They often involve a complex signal correction to compensate for atmospheric reabsorption and require long computing time. In contrast, the exploitation of solar Fraunhofer lines is devoid of atmospheric disturbances. We propose a new retrieval method for red and far-red SIF directly from up-welling radiance spectra in the spectral range between 650 nm and 810 nm by applying Partial Least Squares (PLS) regression machine learning. Solar Fraunhofer lines are exploited for SIF retrieval with the PLS approach by excluding telluric absorption features. The PLS models are trained and tested on synthetic reflectance and SIF data modeled with SCOPE. We identified a logarithmic relationship of the retrieval error with respect to signal-to-noise ratio of the instrument. The approach has been tested with real-world data measured by the Fluorescence Box (FloX), and evaluated against two well-established retrieval methods: the spectral fitting method (SFM) and the singular value decomposition (SVD). PLS models exploiting solar Fraunhofer lines retrieved meaningful SIF values with high precision and demonstrated robustness against atmospheric reabsorption, including from a 100m tall tower. In addition, PLS retrieval requires no complex correction for atmospheric reabsorption and computes 37 times faster than SFM. Hence, PLS retrieval allows fast and robust exploitation of SIF from solar Fraunhofer lines with high precision under conditions in which other retrieval approaches require complex atmospheric correction.

遥感利用太阳诱导的叶绿素荧光 (SIF) 作为从田间到机载和卫星传感器的光合作用的代理。SIF 的调查提供了一种从当地到全球范围内研究植被功能的独特方法。然而，SIF 信号的无源光学检索仍然具有挑战性。常见的反演方法直接从大气氧带中的下涌和上涌辐射中提取 SIF 填充。它们通常涉及复杂的信号校正以补偿大气再吸收，并且需要较长的计算时间。相比之下，太阳能弗劳恩霍夫线的开发没有大气干扰。我们通过应用偏最小二乘 (PLS) 回归机器学习，直接从 650 nm 和 810 nm 之间的光谱范围内的上升辐射光谱中提出了一种新的红色和远红色 SIF 检索方法。通过排除大地吸收特征，利用 PLS 方法利用太阳能弗劳恩霍夫线进行 SIF 检索。PLS 模型在使用 SCOPE 建模的合成反射率和 SIF 数据上进行训练和测试。我们确定了检索误差与仪器信噪比的对数关系。该方法已通过荧光盒 (FloX) 测量的真实数据进行了测试，并针对两种成熟的检索方法进行了评估：光谱拟合方法 (SFM) 和奇异值分解 (SVD)。利用太阳能 Fraunhofer 线的 PLS 模型以高精度检索了有意义的 SIF 值，并证明了对大气再吸收的鲁棒性，包括来自 100m 高的塔。此外，PLS 反演不需要复杂的大气再吸收校正，计算速度比 SFM 快 37 倍。因此，PLS 反演允许在其他反演方法需要复杂大气校正的条件下，以高精度从太阳弗劳恩霍夫线快速、稳健地利用 SIF。