Gross primary production estimation in crops using solely remotely sensed data

Gross primary production (GPP) is a measure for crop productivity, indicating yield and expressing C exchange of agro-eco-systems. A multitude of satellite sensors at varying spatial and spectral resolution brings a possibility to use remotely sensed data for regional and global GPP estimation. More work is still needed to develop algorithms for GPP estimation applicable to multiple, if not all, vegetation types, phenological phases, and environmental conditions. This study employed neural networks (NN), multiple linear regressions (MLR), and vegetation indices (VI) to develop algorithms for GPP estimation based solely on remotely sensed data in two crops, maize (Zea mays L.) and soybean [Gly-cine max (L.) Merr.], with contrasting canopy architectures, leaf structures, and photosynthetic pathways. The focus of the study was to devise algorithms not requiring re-parameterization for different crop species. Data used in the models included in situ hyperspectral reflectance and satellite surface reflectance products. For the tested NN, MLR, and VI algorithms, the bands selected to obtain minimal errors in maize and soybean combined were mainly located in red edge and near infrared (NIR) spectral regions. For both in situ reflectance and satellite surface reflectance, a NN estimated GPP with normalized root mean square errors (NRMSE) below 14 and 18.7%, respectively, and VI using bands in red edge and NIR with NRMSE 15.6 and 20.4%, respectively. The results showed that the models based on the red edge and NIR bands may facilitate accurate assessments of crop GPP at multiple scales, from close range to satellite platforms.