Remotely sensed spatio-temporal trends of irrigation agriculture in northwestern India
Abstract
Irrigated agricultural production plays a key role in covering the world’s food demand. Its importance will grow in the future given increasing population numbers and uncertain climate. Irrigation, however, has also a major impact on water resources, esp. in the drylands on the planet. For example, most of the large-scale problems of aquifer mining can be linked to groundwater-irrigated agriculture. South Asia is one of these regions of concern where roughly 40 percent of the total global groundwater irrigated area is located. In India, almost half of the total agricultural area is irrigated and it is estimated that groundwater irrigation in the country sustains 27 million ha. Esp. in the northwestern part of the country, water tables are falling at increasing rates that give rise to concern about the future viability of irrigation there. Since the majority of food grains in India are produced in that region, this development is a direct threat to the national food security with potentially global implications. We present a novel remote sensing approach to map the temporal development of irrigated agriculture at large spatial scales with high accuracy. We use time series data from the Moderate Resolution Imaging Spectroradiometer (MODIS) on NDVI and surface temperature as well as high-resolution precipitation data from the Indian Meteorological Department from 2000 - 2008 and ancillary data for our supervised classification approach. A cascade of classifiers was chosen to deal with the problem of obtaining labeled examples. A first stage classifier uses large regions of known irrigated and non-irrigated areas to learn a rough estimate of the multi-dimensional time series signature on variables of interest in non-irrigated areas. An estimate of the probability of non-irrigation is generated and passed to a second stage classifier along with the variables used to derive it. The second stage classifier is trained with a small dataset of very high quality estimates of irrigated area, and builds on the output of the first stage to produce a final estimate of the percentage of area irrigated at the pixel scale. The choice of first stage classifier is driven by the need of producing reasonable estimates from large datasets of low-quality data with noise on predictors and labels. The second stage one is chosen with generalization capability in mind, even when small samples are used. We apply our method to northwestern India where our results complement, in a temporal sense, information on irrigation obtained from local census. Due to the linkage of groundwater overdraft and irrigation, the accurate determination of the spatio-temporal evolution of irrigated area coverage allows us to assess the impact of groundwater overdraft and anticipate the consequences of inappropriate water management policies in the region.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2009
- Bibcode:
- 2009AGUFMGC21A0731C
- Keywords:
-
- 0402 BIOGEOSCIENCES / Agricultural systems;
- 1855 HYDROLOGY / Remote sensing;
- 1914 INFORMATICS / Data mining;
- 1942 INFORMATICS / Machine learning