Reference evapotranspiration (ET0) is essential for the management of water resources, particularly for scheduling irrigation and assessing the water needs of crops. However, accurately estimating ET0 is often difficult in economically developing nations due to insufficient availability of climatic data and the reliance on restricted datasets. This research assessed the effectiveness of three machine learning (ML) models-Random Forest (RF), Artificial Neural Networks (ANN), and Long Short-Term Memory (LSTM)-in predicting ET0 employing various input parameters groupings. The study utilized 38 years of meteorological data obtained from the IMD for three districts in Haryana, incorporating inputs such as maximum and minimum temperatures (Tmax, Tmin), relative humidity (RH), wind speed (WS), and solar radiation (SR), with ETo-FAO-56-PM values serving as the target outputs. The performance of the models was evaluated using statistical metrics, including Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), coefficient of determination (R2), and Mean Absolute Percentage Error (MAPE). The results demonstrated that all models achieved accurate ET0 predictions, with the full dataset identified as the optimal input combination. For limited datasets, combinations including temperature, wind speed, and solar radiation were found to be the most effective. In cases of minimal dataset using only temperature data (Tmax, Tmin), RF yielded the best performance during the training phase (R2 = 0.93-0.94). However, during testing, LSTM outperformed RF and ANN across all districts, achieving higher accuracy (R2 approximate to 0.77) and lower errors (MAPE approximate to 17-18 %). Additionally, an uncertainty analysis was conducted to assess the robustness of the models using a Monte Carlo-based approach. The results indicated that LSTM captured extreme ET0 values with broader confidence intervals, reflecting higher sensitivity but lower prediction uncertainties overall, whereas ANN produced tighter intervals with lower variability, and RF offered a balanced performance between accuracy and uncertainty. These findings confirm that LSTM is the most reliable model for ET0 estimation under data-scarce conditions. The temperature based Hargreave-Samani model outperformed the radiation based Priestely-Taylor model in estimation of ET0. The analysis revealed that the LSTM model exhibited lower prediction uncertainties compared to RF and ANN, further highlighting its reliability for ET0 estimation under data-scarce conditions. The results of this research offer a dependable approach for estimating ET0 in semi-arid regions with limited data availability. These findings provide a robust approach for ET0 estimation in semi-arid regions, offering practical guidance for efficient water management and supporting climate-resilient agriculture, thereby contributing to food security in regions with similar agroclimatic conditions to Haryana.