Fig.4. Consequences of biased input data on LUE-NDWI GPP at global scale. (a) the mean annual GPP from 2000 to 2018 after data correction. (b) the difference of mean annual GPP from 2000 to 2018, GPPuncorrected – GPPcorrected. Symbols indicate various levels of difference: ‘–’: -50%-(-30%), ‘-’: -30%-(-10%), ‘±’: -10%-10%, ‘+’: 10%-30%, ‘++’: 30%-50%. The inset bar chart shows the global distribution of the difference proportions. (c) Comparison of annual global GPP estimates from various models. The number after each model’s name in parentheses is the number of years of the model in the interquartile range (IQR), indicated by the shaded region. LUE-NDWI and LUE-NDWI1 are GPP estimated by LUE-NDWI with corrected and uncorrected input data, respectively. (d) Boxplot of annual GPP values during the study period for each model. (e) Trends of annual GPP (Pg C yr−1) by model. Symbols of LUE models are in black, machine learning models in green, biophysical models in blue.