GPP-net: a robust high-resolution GPP estimation network for Sentinel-2 using only surface reflectance and photosynthetically active radiation

High-resolution gross primary productivity (GPP) estimation is crucial for ecological and agricultural applications that require fine spatial details to capture GPP heterogeneity. Satellite-based GPP estimation usually relies on land cover and meteorological data. However, the misclassification of land cover data and coarse resolution of meteorological data greatly increase the uncertainty. Here, we propose a robust high-resolution GPP estimation deep learning (DL) network, named GPP-net, using only satellite surface reflectance (SR) from Sentinel-2 and photosynthetically active radiation (PAR). Specifically, GPP-net is based on a fully 1-D convolutional encoder-decoder network combined with a spectral band importance estimation module. To enhance the generalization of GPP-net, we ran the soil-canopy energy balance radiative transfer (SCOPE) model, and then combined these SCOPE-simulated reflectance data with GPP and PAR data extracted from FLUXNET2015 to pre-train GPP-net. Compared to benchmark models including near-infrared reflectance of vegetation multiplied by incoming sunlight (NIRvP), partial least squares (PLS) and random forest (RF), GPP-net improved half-hourly and daily GPP retrieval across seven plant functional types (PFTs) including four forest types, cropland, grassland and wetland. Owing to its robust nonlinear feature learning capabilities, GPP-net also facilitated robust GPP estimation across both C3 and C4 vegetation. We found that GPP-net could reliably estimate GPP under drought and heatwave conditions, with minimal improvement from including vapor pressure deficit (VPD) as a predictor. Furthermore, GPP-net demonstrated great robustness to soil effects in GPP mapping, and had strong ability in capturing inter-annual variability of GPP. The pretraining paradigm enabled us to fully leverage historical data, and the DL framework ensured that the model generalization continually improves as new data is integrated. Our model dispenses with land cover data and minimizes the requirements of coarse-resolution meteorological data for high-resolution GPP estimation, which could support future efforts in global high-resolution GPP mapping.