FSIF-PCNN: A physics-constrained neural network for full-spectrum solar-induced chlorophyll fluorescence reconstruction

Full-spectrum solar-induced chlorophyll fluorescence (SIF; 650–800 nm) provides a powerful means to characterize vegetation functional status. However, current retrieval methods still exhibit substantial uncertainties due to their reliance on fixed spectral-shape assumptions or strong priors for SIF and reflectance reconstruction, introducing inherent model-form biases when these assumptions mismatch observations. To address this limitation, we propose a physics-constrained neural network, FSIF-PCNN, for flexible and robust full-spectrum SIF reconstruction. FSIF-PCNN leverages a U-Net architecture to extract multi-scale spectral features, while enforcing physical plausibility through radiative-transfer consistency constraints and spectral smoothness regularization. The method was evaluated using simulated test datasets and independent tower-based measurements through comparison with two existing methods (FSIF-iPCA and SIFFI). On the simulated test datasets, FSIF-PCNN achieved minimal systematic bias (ARE = -0.034 %) and superior spectral reconstruction accuracy (RMSE = 0.011 mW m −2 nm −1 sr -1 ; mean RRMSE = 3.87 %), outperforming FSIF-iPCA (mean RRMSE = 13.70 %) and SIFFI (mean RRMSE = 11.98 %). Tower-based validations confirmed exceptional agreement with SFM retrievals in the O 2 -A and O 2 -B bands (site-specific R 2 > 0.95). Compared with FSIF-iPCA and SIFFI, FSIF-PCNN demonstrated stronger accuracy and robustness across all validation sites, achieving site-mean RRMSE values of 5.74 % (DM), 4.46 % (GC), and 9.62 % (XTS) in the O 2 -A band, and 14.64 % (DM), 6.85 % (GC), and 11.22 % (XTS) in the O 2 -B band. These results demonstrate that FSIF-PCNN provides a flexible and physically constrained framework for full-spectrum SIF reconstruction, achieving improved agreement with independent references at the validated red and far-red bands.