Continuous ambient air quality monitoring stations report PM_2.5 at a small number of discrete points, whereas exposure and health-impact studies require concentrations at arbitrary locations. This study identifies the interpolation technique that predicts PM_2.5 over Delhi most accurately and, unlike comparisons that adopt software defaults, optimises the governing parameters of every technique before comparing them. Four techniques available in ArcGIS Geostatistical Analyst were examined: inverse distance weighting (IDW), spline, kriging and cokriging. For each technique three parameters were varied systematically, and leave-one-out cross-validation across 24 monitoring stations gave the root mean square prediction error for every combination, in a spatial phase (10 dates) and a temporal phase (24 stations). Cokriging gave the lowest mean error in both phases, 26.21 and 26.03 µg/ m3 respectively, improving on IDW by 6.9 and 4.3 per cent; wind speed, temperature and relative humidity were selected as the optimal secondary variables in both phases. The margin over the deterministic techniques narrows sharply in the temporal phase. Prediction surfaces generated with the optimised model reproduce the two annual pollution episodes reported for Delhi.

PM_2.5, Spatial interpolation, Cokriging, Cross-validation, ArcGIS.