Deep Learning-Based Prediction Assessment and Statistical Bias Attribution for Summer Terrestrial Carbon Flux

To accurately estimate net ecosystem production (NEP) for regional carbon budget assessments, this study focuses on the "five provinces and one city" in East China. By integrating multi-source remote sensing data, meteorological observations, and model outputs, a spatiotemporal carbon flux prediction model based on Convolutional Neural Networks (CNN) was constructed to reconstruct summer terrestrial NEP and attribute statistical biases. The results indicate that: (1) The MARS-CNN model effectively captures the spatiotemporal characteristics of NEP, with an average prediction error of only 0.3676 gC·m⁻²·day⁻¹, accurately reproducing the "high in the south, low in the north" spatial pattern of summer NEP in East China. (2) The prediction bias of the CNN model primarily stems from the variance effect, where the high-variance group contributes significantly to the overall prediction error. This conclusion was validated against TRENDY multi-model results, suggesting that controlling internal model variance is key to improving regional carbon flux prediction accuracy.