Tabular Generative Adversarial Networks (TabGANs) for Flood Forecasting from Meteorological data

Over the centuries, the flood has turned out to be one of the most devastating disasters, while measures taken to mitigate its effect have been significantly poor. Machine Learning (ML) and Deep Learning (DL) could be used to forecast floods, which could aid the government in more effective flood management and significantly reduce the economic and infrastructural damage caused by floods. However, the lack of proper and latest datasets makes flood forecasting difficult and inaccurate. Moreover, for ML and DL algorithms, an adequate dataset is required. Data augmentation is a common approach in this context. For images, it is very straightforward to do that by using various transformations on the images. On the other hand, random tabular data generation is one of the options, but it has various limitations. Thus, a Generative Adversarial Network (GAN) could be used to generate synthetic data. Typically, GAN is involved in image datasets and is hardly ever explored in tabular meteorological datasets. Hence, this study mainly focuses on applying CTGAN in numerical data augmentation in contrast to random data generation. The result of using CTGAN in contrast to randomly generating data is an increase in the performance of flood-forecasting ML models. The performances of these two approaches were conducted against a series of 13 ML models, k nearest neighbor (KNN), Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Ensemble Learning (EL), Gradient Boosting (GrBT), Adaboost (AdaBT), Catboost (CatBT), Light GBM (LGBM), Support Vector Machine (SVM) kernels, Gaussian Naive Bayes (GNB) and Bernoulli Naive Bayes (BNB) were carefully considered for evaluation.