Prediction of agricultural drought behavior using the Long Short-Term Memory Network (LSTM) in the central area of the Gulf of Mexico

Drought is a slowly evolving and highly complex phenomenon that significantly affects human activities. Innovative approaches are being proposed to predict their future patterns, including using artificial intelligence. In particular, this rapidly growing field addresses issues such as high computational costs and the intricacies involved in managing techniques related to general circulation models. Therefore, our objective was to develop a short-term early detection methodology for agricultural drought based on the Normalized Difference Drought Index (NDDI) and the Long Short-Term Memory Network (LSTM) in the central region of the Gulf of Mexico. The methodology entailed calculating the NDDI values from Landsat 8-9 multispectral images for the period spanning from 2013 to 2024 and subsequently comparing the predictions through seven performance measures. In the case of the LSTM configuration, the Bayesian optimizer was employed to identify the optimal architecture and hyperparameters. The final structure consisted of an input layer, three LSTM layers, three dropout layers, a fully connected layer, and a regression output layer. Furthermore, the LSTM network was trained in four test scenarios (10, 30, 50, and 100 epochs) compared with the Gated Recurrent Unit (GRU) neural network. Finally, the efficacy of the proposed methodology was evaluated using seven distinct metrics. The results show that the LSTM network can predict agricultural drought behavior with the lowest error rates in the evaluation metrics RMSE (0.12 to 0.26), MAE (0.08 to 0.21), and \(d_j\) $d_j$ (0.59 to 0.91).