RESUMEN
Accurately estimating the net ecosystem exchange of CO2 (NEE) in cropland ecosystems is essential for understanding the impacts of agricultural practices and climate conditions. However, significant uncertainties persist in the estimation of regional cropland NEE due to landscape heterogeneity and variations in the efficacy of upscaling models. Here, we applied an integrated approach that combined object-based image analysis (OBIA) techniques with advanced machine learning (ML) approaches to upscale regional cropland NEE. We conducted a thorough evaluation of the upscaling approach across four distinct cropland areas characterized by diverse climate conditions. Our study confirmed that OBIA techniques can efficiently segment cropland objects, thereby enhancing the representation and accuracy of characteristics relevant to cropland features. The sequential least squares programming algorithm, among the three methods used for ML model integration, demonstrated exceptional performance in predicting NEE, with an R2 value exceeding 0.80 across all study areas and peaking at 0.90 in the most successful area. On average, there was an 18 % improvement compared to the poorest-performing ML model and a 6 % enhancement compared to the best-performing ML model. The upscaled regional products exhibited superior performance in characterizing cropland NEE patterns compared to pixel-based products. Additionally, we utilized the SHapley Additive exPlanations (SHAP) to assess driver importance, revealing that phenology and radiation had the greatest influence on prediction accuracy, followed by temperature and soil moisture. This study highlights the potential of integrating OBIA techniques with machine learning approaches for upscaling regional cropland NEE, while concurrently reducing estimation uncertainties.
