Pooled Spatio-Temporal Machine-Learning Forecasting of Ambient CO2 Using IoT-Based Meteorological Sensor Networks: A SHAP-Based Explainability Study Across Heterogeneous Greek Sites

Forecasting ambient CO2 concentration at the local scale remains comparatively under-studied relative to emissions or global background-concentration forecasting, particularly with respect to a direct comparison of pooled (data from multiple sites combined into a single model) versus site-specific (the same stations, modeled independently) strategies over heterogeneous IoT sensor networks. This work addresses that gap by developing and evaluating a unified machine-learning framework on data from ten agrometeorological stations of the Hellenic National Meteorological Service, geographically distributed across the Greek territory and spanning six environmental typologies (coastal, island, urban/industrial, lowland/rural, semi-urban, and mountain/forest environments), using 5 min data from the spring season (1 April–2 June 2026). Five model families—Persistence, ARIMA/SARIMAX, Random Forest, XGBoost, and LightGBM—were evaluated under a strictly chronological data-splitting scheme (40 sliding train/validation pairs, a held-out 20% test set) and a systematic paired statistical testing procedure (Shapiro–Wilk, paired t-test, Wilcoxon signed-rank) for selecting the best hyperparameter combinations, in both pooled and site-specific configurations. The results show that at eight of the ten stations, the simple persistence baseline prevails, owing to the strong autocorrelation of CO2 concentration in stable environments, whereas at the two mountain stations (Vitsi, Parnetha), where CO2 dynamics are inherently more unstable, machine-learning models outperform it (at Vitsi, RMSE 5.73 versus 7.46 ppm and R2 0.84 versus 0.73). SHAP analysis confirms and explains this pattern, revealing a shift of predictive weight away from the lagged CO2 value and toward relative humidity and temperature precisely in the unstable environments. Furthermore, the pooled strategy proved more robust than the site-specific one for every machine-learning model (e.g., pooled Random Forest RMSE = 5.43 versus a mean of 16.20 for the site-specific models), mainly because it is less affected by seasonal level shifts beyond the training range; the pooled machine-learning models did not, however, outperform persistence. For the spring period studied, the findings indicate that the value of artificial intelligence, and especially of explainable artificial intelligence, is not universal but concentrates precisely where a stable, predictable environment is absent—a finding with direct implications for the targeted design of future CO2 monitoring systems.