A Machine Learning Approach to Predicting Groundwater Potential in Zambia from Geologic and Remotely Sensed Variables

Abstract

Rapid agricultural expansion is projected to occur in Zambia in order to meet rapidly rising food production demands in sub-Saharan Africa. Zambia has ample arable land and water resources, yet most Zambian farmers currently rely on rainfed agriculture, which is vulnerable to climatic variability and produces much lower yields than irrigated agriculture. Although groundwater is a largely untapped resource in Zambia, a lack of present knowledge as to the distribution of groundwater potential in the country hinders the development of extensive groundwater irrigation. In this study, I propose a new, empirical approach for predicting groundwater potential that uses the random forests machine learning algorithm to identify the relationship between geologic and remotely sensed variables and existing borehole yields. This approach produces a regression model, which provides numerical estimates of potential yields with 95% prediction intervals, and a classification model, which divides the land into areas of high, medium, and low groundwater potential. These models improve on previous empirical models for predicting groundwater potential by transforming predictor variables into metrics characterizing physical components of the water cycle, eliminating the need for assigning a priori weights to define the importance of each variable, and providing a measurement of uncertainty in the overall accuracy of the model and in each prediction. Using this new approach, I find that approximately 10,681 sq km, or 1.46% of the total land area in Zambia has high groundwater potential. These areas are primarily found in the central, western, and northwestern regions of Zambia and should be targeted by subsequent, local groundwater investigations.