A physically-informed long short-term memory-based tool for predicting extensive droughts in the distant future

Abstract

Agricultural drought is a specific type of drought that impacts agricultural activities and crop yield by lower precipitation and shortages in soil water content. Developing a drought prediction tool is crucial as it can aid farmers and authorities in devising mitigation strategies like crop rotation and deficit irrigation. We developed a long-term, large-scale drought prediction tool solely based on remote-sensing data where drought intensity was measured by an enhanced combined drought index (ECDI) that utilized a weighted summation of four climatic variables: precipitation, temperature, Normalized Differenced Vegetation Index, and soil moisture. The State of Texas in the US is selected as our case study area. We trained a Long-Short Term Memory network with past 21 years of training data to predict the four climatic variables and calculated ECDI for the next 12 months. For model evaluation, we compared results of predicted droughts from ECDI to actual drought events based on SPI-3 (Standardized Precipitation Index with a three-month time scale). Results showed that ECDI and SPI exhibit similar spatial distribution of droughts but with different intensities. We also compared ECDI/SPI values to US Drought Monitor (USDM) maps which show experts’ assessments of conditions related to dryness and drought. ECDI results were similar to USDM in case of drought extent but yielded different intensities. Results of this study showed that remote sensing data can be successfully used to predict future agricultural droughts for a longer period (12 months) and for a large-scale area to assist farmers and policymakers with designing mitigation measures.