Deep Neural Network Augmented by Non-Negative Matrix Factorization: A Unified Approach to Rainfall Prediction

Precognition of short-term rainfall raises two general challenges: coping with highly non-linear and noisy data and applying advanced model methods that possess the capability of modeling complex rainfall patterns. The solution to the challenges above in this study introduces a novel approach through a combination of Deep Learning as a DNN and NMF for short-term prediction of rainfall of flash flood locations of Bangladesh, namely Sylhet and Chittagong. Rainfall measurements were acquired using the Google Earth Engine (GEE) and analyzed using NMF to remove noise from the rainfall by revealing prominent frequency components. In decomposing the data using NMF, the optimum number of components used in model training was set from the investigation of the first and second derivatives of the reconstruction error curve for an optimal level of noise elimination and data preservation. Then a two-hidden-layer Particle Swarm Optimized Deep Neural Network was utilized to learn the rainfall dynamics. The dataset was divided into ratios: 70 (Train):15 (Validation):15 (Test). It was also compared with traditional lag-time-based models. The outcomes demonstrate that the proposed model possessed an R² value of 0.99 for both stations and MSE values of 0.031 for Sylhet and 0.016 for Chittagong, which were much greater than traditional lag-timebased models with substantially smaller R² values of 0.30 and 0.34 and substantially greater MSE values of 212.26 and 132.37, respectively. The results show the effectiveness of NMF-based deep learning in offering accurate rainfall forecasting and realtime flood management.