Droughts have caused millions of deaths since the 1900s and have affected billions of people.
Droughts are considered one of the costliest hazards and have a baneful impact on the economy,
agriculture, environment, and societal development. In order to be able to mitigate the negative
impacts of droughts, we need to shift from the traditional way of providing emergency
assistance to a proactive approach that builds resilience. Drought-forecasting systems play a
crucial role in drought management. The present study proposes an advanced approach based on
deep learning to predict droughts in Indian parts of the Indus River basin at a lead time of ninety
days. A Long Short-Term Memory (LSTM) based deep learning architecture has been used in
the study to improve the drought prediction accuracy. It has been known for a very long time that
ENSO parameters have a significant relationship with Indian drought incidents hence a good
understanding of the role of each variable behind the prediction is the key to build an efficient
forecasting model. The study uses an Explainable AI technique, SHAP (Shapley Additive
Explanations) to understand the effect of various features such as Rainfall, NINO, IOD, SOI and
Temperature values on the predictions. The data was split into two parts, training data (1979-
2010) and testing data (2010-2020). R-Squared error (R2) and Root Mean Square Error (RMSE)
has been used to measure the model’s performance. Results show that LSTM is a promising tool
while predicting droughts in the region and the interpretation of the forecasting model facilitates
comprehension of the effect of climate variables in governing drought dynamics.

R-Squared error (R2) and Root Mean Square Error (RMSE)
has been used to measure the model’s performance. Results show that LSTM is a promising tool
while predicting droughts in the region and the interpretation of the forecasting model facilitates
comprehension of the effect of climate variables in governing drought dynamics.