Integrating Indigenous Knowledge with Deep Learning for Meteorological Drought Prediction

Abstract

Drought is a natural disaster creating huge impacts in three areas of economy, environment and social. It becomes hard to predict drought, its onset and duration due to complex interaction of multiple factors. Though various scales and forecasting methods have been proposed, recent climatic variations caused by Global warming makes most of the scales and forecasting methods inaccurate. Most of existing solutions are based on seasonal behavior and correlation to other influencing factors like temperature and humidity. They are at larger coverage level and not specialized to cover smaller regions. Also with factors like global warming are affecting the baseline periodicity assumptions, there is a need to improvise meteorological factors based drought prediction methods.  This work proposes a solution to this problem by integrating indigenous knowledge (IK) with deep learning forecasting methods through attention mechanism referred as IK fused attention networks. The indigenous knowledge view over precipitation, temperature, wind speed and humidity are integrated with LSTM based forecasting through attention mechanism to improve the accuracy of drought prediction. The performance of the proposed solution was tested against meteorological data collected from Karnataka disaster monitoring center for Chitradurga district of Karnataka. The proposed IK fused attention network is able to provide at least 1.2% higher NSE and 33% lower MAE in prediction of SPI compared to existing works