Drip irrigation in arid and semi-arid regions is frequently compromised by emitter clogging from calcium carbonate scaling, traditionally assessed via the Langelier Saturation Index (LSI) using laboratory-intensive measurements of calcium hardness and alkalinity that preclude real-time monitoring. This study develops a fielddeployable, real-time LSI prediction framework using only three low-cost, continuously measurable sensor inputs: pH, temperature, and electrical conductivity (EC). A 25-year (1991-2015) hydrological dataset comprising 4,633 samples from Khuzestan Province, Iran, was used to train and compare seven optimized models-classical machine learning (GA-tuned SVR, RF, XGBoost), deep learning sequence models (Bayesian-tuned CNN, LSTM, GRU), and a novel hybrid Wavelet-Artificial Neural Network (WANN). Models were evaluated across seven input combinations, with performance assessed via RMSE, MAE, NSE, R2, distributional tests (Kolmogorov-Smirnov), rank correlations (Kendall's tau), bootstrapped 99% confidence intervals, and feature importance analysis. The full three-input scenario (EC + pH + Temp) yielded the highest accuracy, with GA-XGBoost (NSE = 0.844, RMSE = 0.132) and Bayesian-WANN (NSE = 0.838, RMSE = 0.134) outperforming deep learning models. Random Forestbased feature importance revealed pH as the dominant driver (58.8%), followed by EC (29.6%) and temperature (11.5%), explaining the modest NSE gain from including temperature. Bootstrap hypothesis testing confirmed statistical equivalence among top performers (GA-XGBoost, GA-SVR, GA-RF, B-WANN). The EC + pH pairing proved a robust alternative (NSE approximate to 0.79-0.80) when temperature data are unavailable. By enabling proactive, sensor-driven scaling risk assessment on lightweight edge devices, this framework overcomes limitations of conventional equilibrium-based indices, offering a practical tool for clogging prevention and sustainable water management in resource-constrained agriculture.