Towards Generalizable Groundwater Withdrawal Predictions: How Much Data Do We Need?
Abstract
The heavy reliance of regions with arid to semi-arid climates on groundwater for irrigation and other uses has caused widespread aquifer depletion in many regions. Water management decisions are often made with scarce data, and one of the most important datasets for these decisions, groundwater withdrawals, is seldom known with accuracy. Recent methods employing remote sensing data to estimate groundwater storage trends in machine learning frameworks have produced promising results. Typically, these models were trained and tested in areas with rich groundwater datasets, or less commonly in regions with sparse datasets where testing the model accuracy is challenging. The proposed study assesses the quantity and types of data required to build a parsimonious and generalizable groundwater withdrawal machine learning prediction model at a local scale using the random forest algorithm. The study is undertaken in Northwest Kansas Groundwater Management District No. 4 in the central United States. As the availability of in-situ data is extremely limited globally, machine learning models constructed in regions where in-situ data are available can be used to study similar regions where data are limited. A combination of in-situ (groundwater withdrawal rates), remote sensing (precipitation and evapotranspiration) along with estimated crop water demand are used as predictors for the construction of the machine learning model. These data are selected by studying the major factors that drive groundwater pumping in the study area from 2008 to 2021. All the data are converted to raster format with the same grid size of 2 km by 2 km. Crop water demand is mapped to the cropland data layer map obtained from the United States Department of Agriculture. The preliminary results indicate that climate (precipitation) and land use (major crop types - primarily corn) are the major drivers of water demand in the area.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H25T1364A