Through more than 400 wells, tritium leakage of Pickering Nuclear Generating Station has been monitored seasonally since 1999. Sampling and maintenance of monitoring wells being costly, it is required limiting number of samples while ensuring monitoring objectives. This study aims at proposing a geostatistical approach for sample reduction while meeting the monitoring objectives. So, four objectives were defined: (i) Geographical coverage. (ii) Denser sampling where tritium variability is high. (iii) Delimiting the threshold of 300 kBq/L, also (iv) the cut-off of 3000 kBq/L. These objectives were quantified using geostatistical measures, served as cost functions in heuristic optimization algorithms, implemented using scripting capacity of Isatis.neo software. The algorithm was successful in sampling optimization of 2010Q4 (fourth season) according to the first, third and fourth measures. The first measure selects spatially evenly distributed wells, necessary for unknown leakages. The third and the fourth geostatistical measures suggest sampling around a decisive tritium concentration. Considering these objectives, 22 samples were removed (totally 50 samples) while deterioration of characterization accuracy is negligible. Temporal variogram revealed tritium correlation over three years. So previously acquired samples could be used to improve the monitoring in scarcely sampled areas. But sensitivity analysis showed that samples older than four seasons do not improve the current season contamination characterization. The conditional optimization was applied to samples of 2010 (all seasons). Previous samples improved the first and the last geostatistical measure, while deteriorated the second and the third measures.