Creating a drought prediction model using convolutional neural networks

Droughts kill over 45,000 people yearly worldwide, with climate change likely to worsen these effects. Despite this, researchers have struggled to develop a method that accurately predicts the location of droughts. One of the most recent and accurate drought prediction models is DroughtCast. DroughtCast utilizes a neural network along with precipitation, temperature, and other weather data to predict the United States Drought Monitor (USDM) index of a given week; however, this model does not consider the contextual aspect of weather forecasting. Predicting weather exclusively using data from a single location will never be as successful as predicting the weather using data from that point in addition to its surroundings. As a result, we created a novel Convolutional Neural Network (CNN) based upon the U-Net architecture to predict future USDM indices by using the current USDM index, historical USDM indices, and a 10-year (2010–2019) dataset containing weather data such as precipitation and Snow Water Equivalent that was obtained from DAYMET, a NASA database for weather across all North America. While the data utilized in our model is similar to DroughtCast’s data, the model architectures are different. We hypothesized that this new architecture would improve the accuracy of our prediction. In comparison to DroughtCast, the mean- squared-error of the CNN Model dropped by 85%, 98%, and 97% for prediction times of 1 week, 6 weeks, and 12 weeks respectively, meaning that a vastly more accurate prediction.