The Decision Tree Aided Neuro-Fuzzy Inference Characterization of the Stochastic Hydrology of the Tana Alluvial Aquifer

Abstract

The Tana Alluvial Aquifer is the name given to the little-understood aquifer which is active in the areas bordering the River Tana Flow course as the river weaves its way through the sedimentary plains of Balambala, Garissa, Fafi and Ijara and, finally, into the Tana Delta areas, with the common denominator being the proximity to the Lower Tana catchment, especially the riparian corridor of the River itself, and beyond. The aquifer may extend to between five to fifteen kilometers away from the river channels course way, and at times, it may be felt even 20 kilometers away. The geology of the locality is heterogeneous and comprise sediments whose soil mechanics may not be easily deciphered, since some areas close to the river have very fresh water while others are saline (Bura East in Fafi Sub County easily comes to mind here). There are areas far from the river but bearing fresh water (Mulanjo comes to mind). In some areas, sites close to the river discharge low yield figures, whereas those located farther afield discharge favorably. The water quality and discharge are therefore stochastic variables, subject to chance occurrence. In view of this inconsistency, and on the account of data scarcity, the neuro-fuzzy inference algorithm was developed to map the Universe of Discourse of the Tana Alluvial Aquifer, aka the T.A.A., as it relates to the longitudes, latitudes, depths, and discharges of the aquifers in the study area. The mapping was with respect to aquifer discharge, the variable used to characterize an aquifer, in terms of Transmissivity and Hydraulic Conductivity, thereby defining aquifer recharge propensity. Membership functions were developed using the trapezoidal membership family, and fuzzy rules were appropriately evolved from the fuzzified aquifer data, before finally employing the Sugeno inference engines (in Python) to make predictions of discharge, at each of the T.A.A. aquifer subsets mapped for fresh, saline, hard and brackish water species. The accuracy in the outputs achieved in the areas mapped vindicated the power of the neuro-fuzzy inference systems, as the accuracy oscillated between 92 and 99 percent, when the discharge values predicted were compared with the actual known discharge values of the wells mapped. The water quality class characterization was then undertaken using the decision tree (DT) algorithm in python which gave rise to a 100 percent prediction accuracy. The same DT algorithm could not successfully predict the discrete values of aquifer discharge or EC values, with as much accuracy (but performed excellently with salinity class data), and that was why fuzzy logic was employed. The study vindicated the use of the DT and Fuzzy Logic Algorithms as simple, yet powerful analytical tools, in characterizing the Stochastic Hydrology of the Tana Alluvial Aquifer.