GPP estimation by transfer learning with combined solar-induced chlorophyll fluorescence and eddy covariance data

Gross primary productivity (GPP) plays a crucial role in the carbon exchange between the atmosphere and terrestrial ecosystems. Eddy covariance (EC) method can obtain accurate GPP at the site level, but the sparse distribution limits representativeness. Satellite solar-induced chlorophyll fluorescence (SIF) serves as emerging data of large-scale GPP, yet there are still limitations in its conversion to GPP and spatiotemporal coverage. This study proposes a transfer learning (SIFEC-TL) method to estimate long-term global GPP with high accuracy by combining constraints from SIF and EC data. SIF data are taken as the source domain that provides the spatial information for pre-training, and EC GPP in the target domain provides precise GPP for the machine learning model fine-tuning. To verify the performance of SIFEC-TL, the results are compared with those from machine learning models that use only SIF or EC GPP alone (SIFML and ECML). The results indicate that the SIFEC-TL model demonstrates stronger spatial scalability compared to the SIFML and ECML models, with R2 increasing by 0.132 and 0.036. The SIFEC-TL more effectively captures inter-annual GPP dynamics with underestimation/ overestimation over high/low values in the SIFML and ECML models being well corrected. Furthermore, three different SIF-based GPP are also used as source domains, and the results showed that they only affect pre-training but the final accuracy after fine-tuning remains similar, which indicates SIFEC-TL can obtain stable GPP estimation accuracy regardless of the spatiotemporal coverage of SIF data and its conversion to GPP.