<p>Accurate emission estimates are essential, as governments worldwide report annual greenhouse gas (GHG) inventories to monitor progress toward climate targets under international requirements. However, accurate GHG accounting in wastewater treatment plants (WWTPs) across the Global South is hindered by severe data scarcity—missingness patterns vary by facility and parameter, with essential variables for carbon accounting—including energy use and nitrogen loads—frequently unavailable (missing rates: 80–100% in multiple plants). Conventional machine learning (ML) approaches that rely on algorithmic handling of incomplete data yield misleading predictions and physically implausible insights. We address this gap by integrating probabilistic principal component analysis (PCA) for multivariate imputation with interpretable ML models—Random Forest (RF) and eXtreme Gradient Boosting (XGBoost)—to predict annual CO₂-equivalent emissions across six WWTPs in Santiago de Querétaro, Mexico (2013–2024). Our framework uniquely integrates direct emissions (from on-site CH₄ and N₂O production and indirect CO₂ emissions from grid electricity) with avoided emissions—the GHGs prevented through organic load removal—which together enable a net carbon footprint assessment rarely implemented in data-limited settings. Results show that imputation quality—not just model choice—determines adequate validity; PCA-based imputation increased usable observations, corrected spurious variable rankings, and revealed treated flow and grid electricity as the dominant emission drivers—consistent with engineering principles. XGBoost achieved exceptional in-sample accuracy, yet temporal validation exposed its forecasting fragility, while RF demonstrated greater stability under real-world distributional shifts. Critically, we demonstrate that iterative PCA preserves data geometry better than Multiple Imputation by Chained Equations (MICE), avoiding artificial inflation of correlations in highly collinear systems. This study establishes that rigorous missing-data preprocessing is foundational—not optional—for credible, actionable GHG estimation in resource-constrained urban water systems. The proposed framework offers a scalable, transferable blueprint for dynamic carbon accounting and targeted mitigation in the Global South.</p> Graphical Abstract <p></p> <p>Based on the graphical snapshot, this study was conducted to address the critical challenge of estimating GHG emissions from WWTPs in data-scarce regions of the Global South. The visual summary captures the entire research workflow, beginning with the data source: six WWTPs in Santiago de Querétaro, Mexico, characterized by severe missingness in key parameters like energy consumption and nitrogen loads. The analytical approach is visually segmented into two parallel paths: one using a non-imputation method (algorithmic handling of missing data) and the other employing a probabilistic PCA imputation method. This comparison highlights that conventional Machine Learning (ML) approaches yield misleading predictions, while PCA-based imputation preserves the latent multivariate structure of the system, enabling robust model development. The model application section depicts the use of two interpretable machine learning algorithms —RF and XGBoost— to predict annual CO₂-equivalent emissions. The results are presented through comparative radar plots and time-series graphs, revealing that treated flow and grid electricity are the dominant emission drivers, consistent with engineering principles. Crucially, the graphical abstract emphasizes that the choice of imputation method fundamentally shapes variable importance rankings: improper handling distorts them, whereas PCA restores physical coherence. Finally, the conclusion panel summarizes the framework’s value as a scalable, transferable blueprint for dynamic carbon accounting in resource-constrained settings, supporting climate-resilient sanitation planning under real-world data constraints.</p>