Performance-based assessment of gross primary production(GPP)products in a typical inland river basin of northwestern ChinaOA

Accurate estimation of gross primary production(GPP)is crucial for understanding terrestrial carbon cycling,yet the regional performance of existing GPP products remains insufficiently quantified.This study evaluated four widely used GPP products,i.e.,the Moderate Resolution Imaging Spectroradiometer(MODIS,e.g.,MOD17),Global OCO-2-based Solar-Induced Chlorophyll Fluorescence(SIF)product(GOSIF),Global Land Surface Satellite(GLASS),and Penman-Monteith-Leuning Version 2(PML_V2),across five representative ecosystems in the Heihe River Basin(HRB),northwestern China using 18 eddy covariance(EC)sites during 2007–2022.Multi-scale validation revealed pronounced spatial and ecosystem-dependent differences.Although all products captured the general basin-scale gradient,an analysis of the spatial coefficient of variation(CV)revealed distinct differences in their ability to resolve spatial heterogeneity:PML_V2 and GLASS reasonably captured the observed spatial variability,whereas MOD17 and GOSIF tended to smooth over fine-scale details.Furthermore,a systematic compression of the productivity gradient was evident across products,characterized by considerable underestimation in high-productivity ecosystems(forest land and cropland)and general overestimation in grassland.Temporally,all products performed more reliably in capturing seasonal dynamics than in reproducing inter-annual variations.At the growing season scale,GLASS and GOSIF achieved the highest explanatory power(r>0.95 at several sites),whereas MOD17 exhibited the lowest error(root mean square error(RMSE)=25.26 g C/m^(2) at the Jingyangling site(JYL)).However,inter-annual performance declined markedly,with MOD17 showing weak correlations(r<0.30)at most sites.Ecosystem-specific results identified GOSIF as superior for cropland and wetland ecosystems,while PML_V2 offered the best performance in grassland and desert ecosystems by minimizing systematic bias.Notably,all products consistently failed to establish meaningful correlations(r<0.21)with observations in desert areas due to the low ratios of signal to noise.Consequently,we recommend an ecosystem-dependent application strategy—specifically prioritizing GOSIF for cropland and wetland and PML_V2 for grassland—and urge extreme caution when applying single remote-sensing GPP products in arid desert areas.