Mapping of large irrigated areas in Central Asia using MODIS time series

Remote sensing offers the opportunity to produce land cover classifications for large and remote areas on a yearly basis and is an important tool in regions that lake these information. However, often training and validation data to generate annual land cover maps are not available in necessary quantity - being from one year only or covering only a small extent of the region of interest. This study was focused on land use classifications at regional scale with a special emphasize on annual updates under the constraint of limited sampling data. Often, sampling is reduced to one year or to an unrepresentative area extend within the region of interest. The investigations for the period between 2004 and 2009 were conducted in the irrigation systems of the Amu Darya Delta in Central Asia, where reliable information on crop rotations is required for sustainable land and water management. Annual training and validation data were extracted from high resolution land use classifications. For classification, statistical features based on MODIS time series of vegetation indices, reflectance and land surface temperature (LST) were calculated and a random forest algorithm was applied. By a combination of training data from different years, the accuracy could be enhanced from an overall accuracy of 70% to more than 90% for a focused subregion and also good consistency with high resolution images for the other parts of the data, which has to be confirmed using quantitative validation. A combination of a different number of years was tested. Already two years can be sufficient to generated a robust and transferable random forest to produce yearly land use maps. The study shows the possibility to combine training data from different years for the annual classification of irrigated croplands on a regional scale.