Knowledge extraction from trained ANN drought classification model

Vikas Kumar Vidyarthi

doi:10.1016/j.jhydrol.2020.124804

2020-06-01

PRE

Vikas Kumar Vidyarthi *

Ashu Jain

DOI：10.1016/j.jhydrol.2020.124804

原文链接

原文求助

摘要

En 中文

Artificial neural networks (ANNs) are one of the most widely used techniques for solving a variety of problems including classification and function approximation type of problems. However, the use of ANNs in hydrology is limited to academics and research; and hydrologists/policymakers are skeptical to use the ANNs in operational hydrology, especially where a premium is placed on the comprehensibility of the systems. This is probably because of the perceived 'black-box' nature of the ANNs as they do not provide an insight into their learning process or elucidation of obtaining a particular result. Since a long time, researchers have been trying to improve the comprehensibility of trained ANN models by knowledge extraction for improving their reliability and acceptability among the decision makers for operation purposes. There are few such attempts in hydrology particularly for ANN models developed for function approximation problems such as flow forecasting and evaporation estimation. In this paper, a novel approach is proposed that is capable of extracting knowledge from trained ANN models for classification type of problems in hydrology. This is achieved by extracting knowledge in terms of rules by inducing a Decision Tree using input-output relation of a trained ANN model for forecasting drought classes with standardized precipitation index for Indian Meteorological sub-divisions. The reported researches on knowledge extraction in hydrology is primarily been for the function approximation type of problems, however, the present approach is fundamentally different from them and useful for knowledge extraction exclusively for classification problems in hydrology. The findings of this research indicate that definite rules are learned by ANN models during their training to forecast a specific drought class. These extracted rules are simple, easily implementable, and can be used as the drought class forecasting tool in a drought management activity. The results suggest that due to the nature of the drought index used, the rules extracted from one region may be suitable for drought monitoring of other regions.

Keyword:

Drought Classification

Standard Precipitation Index

Artificial Neural Network

Decision Tree

Knowledge Extraction

Water Resources