Irrigation Water Demand Forecasting
M. A. Khan, Md Z. Islam, M. Hafeez
Background
Water is a precious resource. Agriculture uses about 65% of all the water consumed in Australia per annum, most of which is used for irrigation (Source: Australian Bureau of Statistics). As water demand rises in the face of an increasing population and the effects of climate change, improving irrigation efficiency has become a priority for the Australian government.
Currently a farmer must estimate his/her future water requirement for the crop and make a request to the relevant authority. It then takes seven days for the water to be delivered from upstream. Managing available water and future water requirement is difficult.
The Challenge
This study aims to improve water management practices and maximise water productivity by developing a model for irrigation water demand forecasting. Hence irrigation water can be delivered efficiently based on the crop type and cropping stage, and weather forecast.
Method
This study uses historical water data usage, weather stations meteorological data and information about crops and soil types from spatial data for the Coleambally Irrigation Area in New South Wales. The area contains approximately 79,000 ha of intensive irrigation area out of 400,000 ha, supplying water to 491 irrigation farms (Coleambally Irrigation Co-operative Limited (CICL): http://new.colyirr.com.au - 2015).
Coleambally Irrigation Area in the Murrumbidgee Catchment
Charles Sturt researchers used several data sources to build a web-based forecast tool. One difficulty was to determine the daily water usage. Rather than take an average (water delivered divide by the number of days between water deliveries), the daily water usage was estimated by taking into account water losses via evapotranspiration which included soil evaporation and plant transpiration.
Using a Charles Sturt-developed machine-learning technique based on decision tree learning called Sysfor, the data was then classified and analysed. Other known machine-learning techniques were used as well on these data to evaluate the performance of Sysfor.
Results
The results predicted using the different techniques were compared to the water actually consumed during the summer season of 2008-2009. Overall Sysfor performed well, with an accuracy of 95%, compared to the other techniques used. However for three areas, the water usage predicted by all the models was significantly higher than the actual water usage.
The decision tree learning technique was included in the web-based decision support tool called Coleambally Integrated River Information System (Coleambally IRIS). This tool allows farmers to predict their water requirement seven days in advance.
Irrigation water demand forecast tool
Future Work
More work needs to be done using additional data such as seepage and, soil moisture as well as the the influence of cropping stage on crop water use. Such data can be added to improve the forecast capabilities of Sysfor. The water usage predictions then need to be compared to actual end-user usage.
Selected Publications
- Khan, MA, Islam, Z, Hafeez, M, "Evaluating the Performance of Several Data Mining Methods for Predicting Irrigation Water Requirement", Proc. of the 10th Australasian Data Mining Conference (AusDM 12), Vol. 134, Sydney, December 2012.
- Khan, MA, Islam, Z, Hafeez, M, "Irrigation water demand forecasting: a data pre-processing and data mining approach based on spatio-temporal data", Proceedings of the 9th Australasian Data Mining Conference, Vol. 121, 183-194, Ballarat, 2011.
- Khan, MA, Islam, Z, Hafeez, M, "Irrigation Water Requirement Prediction through Various Data Mining Techniques Applied on a Carefully Pre-processed Dataset", Journal of Research and Practice in Information Technology, Vol. 43, No. 22, May 2011.
- Islam, Md. Z. and Giggins, H., "Knowledge Discovery through SysFor - a Systematically Developed Forest of Multiple Decision Trees". In Proc. Australasian Data Mining Conference (AusDM 11) Ballarat, Australia. CRPIT, Vol. 121, 195-204, 2011
Contacts
- M. A. Khan, Md Z. Islam, M. Hafeez
- School of Computing and Mathematics, Charles Sturt University