Understanding the infiltration dynamics under varying rice cultivation practices is essential for sustainable water resource management, particularly in semi-arid regions with sandy loam soils. This study evaluated soil infiltration behaviour under four rice establishment methods: Direct Seeded Rice with Zero Tillage (DSR-ZT), Conventional Tillage (DSR-CT), Reduced Tillage (DSR-RT), and Puddled Transplanted Rice (PTR). Field experiments were conducted using a double-ring infiltrometer, and infiltration data were modelled using three empirical equations (Kostiakov, Philip, and Horton) alongside two machine learning algorithms (Random Forest (RF) and Support Vector Machine (SVM). The Horton model demonstrated superior consistency among the empirical methods, with R2 values approaching 0.98, and lower RMSE and MAE scores. Among the machine learning techniques, the Random Forest model outperformed SVM in most land-use practices, achieving high NSE values (>= 0.91) and minimal prediction bias. The highest infiltration rates were observed under DSR-ZT owing to the preserved soil structure, whereas PTR significantly reduced the infiltration because of the compacted soil layers. These findings highlight the effectiveness of zero-tillage DSR in enhancing infiltration and groundwater recharge and underscore the potential of ensemble-based machine learning models for accurately predicting infiltration under diverse field conditions. The results support adopting conservation agriculture practices and data-driven modelling tools for informed irrigation planning and soil-water management in rice-growing regions.