In Bangladesh, arsenic in irrigation water is thought to be a major pollutant of the soil-plant system, and an additional public health hazard along with high-arsenic drinking water. Geostatistical methods for describing spatial distributions of arsenic in the water-soil-plant system are an integral component for risk assessment of this contaminant. The purpose of this study was to characterize the spatial pattern of arsenic in ground water, paddy soil and rice grain in Tala Upazilla, an arsenic contaminated administrative unit with an area of 34,000 ha. We collected ground water from 161 selected shallow tubewells (STW) used for irrigation of rice. Paddy soil (0-15 cm depth) and rice samples were collected from a 1 m2 area at each STW site. The water, soil and grain samples were analyzed for total arsenic. Ordinary kriging was used to predict arsenic content in water, soil and rice grain with a 168 x 158 grid dimension from 107 sampling points. Map quality was evaluated by using the remaining 54 sampling points as a validated data set and by cross validation with replacement. We used gstat geostatistical package in the R environment for all geostatistical analysis and visualization. The distributions of water, soil and grain arsenic contents were slightly skewed (skewness <1), but medians were close to the means. Grain arsenic showed higher skewness relative to water and soil arsenic values. There was no effect of agro ecological zones (AEZ) on average arsenic content in groundwater or rice grain. However, soil arsenic was higher in the lower lying Ganges Tidal Flood Plain AEZ than in the High Ganges River Flood Plain AEZ. An exponential model was fitted successfully to variograms for all variables. The variogram of water and grain arsenic indicated large scale spatial correlation with a low nugget variance. In contrast, the variogram of soil arsenic content showed small scale spatial correlation. Predictions of water arsenic were better than those for soil and grain arsenic. Bias values for water aresenic prediction were almost zero and the root mean square error (RSME) values were very low for both validation methods. The mean squared deviation ratio (MSDR) of the prediction was almost 1, meaning that the predictions of water arsenic were very good. Predictions of soil and grain arsenic were poor. Soil and grain arsenic may be influenced by local management choices, such as water mangement and rice variety.