| 24 Jun 2026
Physical Climate Drivers of East Africa’s March-April-May (MAM) seasonal rainfall Identified through Machine Learning Analysis
Abstract. East African March–May rainfall (MAM) remains difficult to predict despite its importance for agriculture, water resources, and disaster preparedness. This study identifies pre-season physical drivers of MAM rainfall and tests their value for probabilistic seasonal prediction. Predictor basins were derived from December and January sea surface temperature (SST), 2 m air temperature (T2), and sea-level pressure (SLP) anomalies relative to 1991–2020, using correlations with the leading mode of East African MAM rainfall and subsequent SHAP-based feature selection. The selected basin-derived indices were applied in Random Forest (RF) and Extreme Gradient Boosting (XGB) models. The dominant predictors appear to be the southern Indian Ocean T2 tendency, Australian and Eurasian T2 gradients, South Pacific and Antarctic T2 signals, Atlantic Niño tendency, and the Euro–African SLP gradient. T2-related predictors dominate both the January and December initialisations, showing that near-surface thermal gradients provide useful information in addition to SST memory. Walker-circulation diagnostics show that these drivers influence rainfall through pressure-gradient changes, tropical overturning, and upper-level wave-train development. For January initialisation, RF and XGB achieve spatially averaged Brier Skill Scores of 0.48 and 0.41, respectively, while the corresponding Area Under the Receiver Operating Characteristic Curve values amounting to 0.72 and 0.65. These results demonstrate that physically constrained machine learning provides promising probabilistic skill for East African MAM rainfall prediction.
Overall assessment
This manuscript addresses an important problem in East African seasonal rainfall prediction by combining physically based climate predictors with explainable machine-learning methods. The study is promising and potentially useful for seasonal forecasting, but several methodological and interpretive issues need clarification before the reported skill and physical conclusions can be considered robust.
Major contributions and achievements
The study provides a useful integration of climate dynamics, SHAP-based feature selection, and RF/XGB probabilistic forecasting. The identification of T2-based predictors in addition to SST signals is particularly interesting, while the circulation diagnostics provide a useful attempt to explain the physical pathways behind the statistical relationships. The reported BSS and AUC values also indicate potentially useful predictive skill.
Major comments
Predictor selection and information leakage: The manuscript should clearly explain whether correlation-based basin identification and SHAP feature selection were performed independently within each training fold. If the full dataset was used before validation, information leakage could inflate the reported forecast skill. A nested or strictly out-of-sample selection procedure is recommended.
Physical drivers versus statistical predictors: The terms “physical controls,” “drivers,” and “influence” appear stronger than supported by correlation and ML methods alone. The authors should distinguish predictive association from causality and moderate the wording unless stronger dynamical evidence is provided.
Justification of EOF1/PC1: The choice of rainfall PC1 as the predictand needs stronger justification. The authors should report its explained variance and spatial pattern and explain why it is preferable to a regional-mean rainfall index. A sensitivity test using regional mean MAM rainfall would strengthen the analysis.
Multiple testing: The use of p < 0.1 for global grid-point correlation screening may produce false-positive regions because many correlations are tested. The authors should address multiple comparisons and provide sensitivity tests using alternative significance thresholds or field-significance methods.
Additional value of T2 predictors: The conclusion that T2 provides information beyond SST should be demonstrated quantitatively. Comparisons of SST-only, T2-only, SLP-only, and combined models would show whether T2 predictors provide genuine incremental forecast skill.
Probabilistic skill assessment: The reported BSS and AUC values are promising, but the reference forecast, forecast categories, sample size, validation procedure, and uncertainty estimates should be clearly described. The meaning of “spatially averaged” BSS should also be clarified given that PC1 is described as the target.
Predictor robustness: The stability of the selected predictors across different training samples should be assessed. Reporting predictor-selection frequency or bootstrap stability would help establish whether the identified climate signals are robust rather than sample-dependent.
Physical mechanism: The proposed SST–T2–SLP–circulation pathway is plausible, but the current analysis appears mainly associative. Additional lagged relationships, circulation composites, moisture convergence, or moisture-budget diagnostics would strengthen the physical interpretation.
Operational relevance: The potential application to ICPAC and other climate services is valuable, but operational readiness should not be implied without independent hindcasts or real-time testing. The wording should be moderated to indicate potential operational usefulness.
Overall recommendation
The manuscript has a strong and relevant research focus, with promising contributions from physically informed ML and the identification of T2-related predictors. However, the reported skill and physical interpretations require stronger validation and clearer methodological documentation. I recommend major revision, particularly addressing information leakage, predictor robustness, incremental T2 skill, and the distinction between predictive association and causal physical drivers.