An Integrated System to Forecast PM10 Concentrations in an Urban Area, Using MODIS Satellite Data

Abstract

An integrated forecasting system, using artificial neural networks (ANNs) and cokriging technique, has been developed to forecast daily mean PM10 concentrations 2 days in advance. The test case has been performed over Milan metropolitan area in northern Italy where PM10 concentrations are continuously monitored by 14 monitoring stations. In the first step, ANNs are identified for each monitoring stations to forecast and in the second step, the forecasted values are interpolated using cokriging algorithms over the whole domain. The use of the MODIS derived PM concentration maps as a secondary variable for cokriging, account for the local spatial patterns of PM10 where no measurements are available. The results are validated in terms of statistical and forecast exceedance indexes. The validation has been performed in two steps: first, the Neural Network model performances have been investigated comparing the point forecast with observations, and then the cokriging algorithm has been validated using leave one out cross validation method. The validation results show good agreement in terms of statistical indexes. The proposed forecasting methodology represents a fast and reliable way to provide decision makers and general public with PM10 forecast over an urban area.