Worldwide food production is under ever-increasing demand. Meanwhile, climate change is disrupting rainfall and evaporation patterns, making agriculture freshwater supplies more uncertain. IPCC models predict an increased variability in rainfall and temperature over most of the globe under climate change. Yet, the effects of climate variability on water security remain poorly resolved. Here we used satellite images and deep-learning convolutional neural networks to analyse the impacts of annual averages, seasonality, climate anomaly, and temporal autocorrelation (or climate reddening) for rain and temperature on the water levels of >100,000 Australian farm dams across 55 years. We found that the risk of empty farm dams increased with warmer annual temperatures, lower yearly rainfall, stronger seasonality, reduced climate anomalies, and higher temporal autocorrelation. We used this information to develop a predictive model and estimate the likelihood of water limitations in farm dams between 1965 and 2050 using historical data and Coupled Model Intercomparison Project Phase 5 (CMIP5) at two climate change scenarios. Results showed that the frequency of empty water reserves has increased 2.5-fold since 1965 and will continue to increase across most (91%) of Australia. We estimated a 37% decline in rural areas with year-round water supplies between 1965 (457,076 km2) and 2050 (285,998 km2). Our continental-scale assessment documents complex temporal and spatial impacts of climate change on agricultural water security, with ramifications for society, economy, and the environment.