Fuzzy Logic and Machine Learning Approaches for Modeling Monthly Rainfall from Meteorological Variables in Sulaymaniyah, Kurdistan Region of Iraq

Monthly rainfall information is essential for water management and rainfed agriculture in semi-arid Kurdistan Region of Iraq. This study developed a Mamdani fuzzy inference system (FIS-RH-T-E) in MATLAB that estimates monthly rainfall in Sulaymaniyah from relative humidity (RH), temperature, and evaporation, using 140 expert rules and data from Bakrajo station (2010–2024, n = 180). The FIS was compared with multiple linear regressions (MLR), random forest, gradient boosting, support vector regression, a multilayer perceptron and monthly climatology, evaluated by leave-one-year-out cross-validation and a 2020–2024 hold-out, including one-month-ahead prediction. Climatology alone achieved a Nash-Sutcliffe efficiency (NSE) of 0.45. The FIS reached a coefficient of determination (R 2 ) of 0.63 and NSE = 0.63 (Kling-Gupta efficiency 0.71), second only to MLR (R 2 = 0.67, NSE = 0.66) and ahead of support vector regression, gradient boosting and random forest (NSE 0.61–0.62); differences among these five models were not significant, whereas all clearly outperformed climatology and multilayer perceptron (NSE = 0.44). Scored on their training data, the machine-learning models reached NSE up to 0.84, which overstates their accuracy. RH dominated both FIS response surfaces and data-driven analyses (anomaly correlation with rainfall 0.63, against − 0.21 for temperature and evaporation). All models underestimated the wettest months; the FIS never exceeded 221 mm, whereas observed rainfall reached 316 mm. No model predicted rainfall one month ahead better than climatology. An expert-defined FIS therefore provides transparent estimates as accurate as machine learning for filling gaps in rainfall records, whereas forecasting requires large-scale predictors.