Mapping Irrigated Areas in China From Remote Sensing and Statistical Data

Spatial information on irrigation is needed for a variety of applications, such as studies on water exchange between the land surface and atmosphere, climate change, and irrigation water requirements, water resources management, hydrological modeling, and agricultural planning. However, it is hard to map irrigated areas automatically by traditional image classification methods because of the high spectral similarity between the same crops with and without irrigation. In this study, we developed three irrigation potential indices by using the time series normalized difference vegetation index (NDVI) and precipitation data. Using these indices and a spatial allocation model, we downscaled the census data on irrigation from administrative units to individual pixels and produced a new irrigation map of China around the year 2000. We collected 614 reference samples (262 irrigated samples and 352 nonirrigated) in mainland China to validate our new irrigation map and also two global irrigation maps: one is produced by the Food and Agriculture Organization of the United Nations and the University of Frankfurt (FAO/UF map), whereas the other is produced by the International Water Management Institute (IWMI map). The overall accuracies of IWMI map (0.0089282°) and the new map (1 km) are 60.91% and 68.40%, respectively. We also resampled the IWMI map and the new map to match the spatial resolution of FAO/UF map (0.0833333°), and calculated the producer accuracies of FAO/UF map, resampled IWMI map, and resampled new irrigation map. The accuracies are 83.2%, 83.2%, and 87.0%, respectively. We further compared the three maps using cluster and outlier analysis and spot analysis. Comparison results suggest that our new map agrees very well with the patterns of irrigated area distribution from the FAO/UF map, but differs greatly from the IWMI map. Results from this study suggest that our method is a promising tool for mapping irrigated areas. It has several advantag- s. First, its inputs are quite simple, and no training samples are needed. Second, our model is general and repeatable. Third, it can be used to map historical irrigated areas. The limitations of our method are also discussed.