Optimization Techniques in ANN for Assessing Meteorological Droughts in Dhaka and Chittagong, Bangladesh

Bangladesh significantly depends on seasonal and annual rainfall for agriculture. Yet, there are flash or regional floods and droughts affecting human-lives, properties, and crops that urge low-cost but accurate warning systems in meeting sustainable developments. The paper aims to predict rainfall in Bangladesh by incorporating the Artificial Neural Network with optimizations based on climatology. The study-site is focused on metropolitans: Dhaka (capital) and Chittagong (port) cities and compared with historical evolutions. The experiments include data mining and a statistical approach for trend analyses before modeling. The observation data, i.e., 24- hr accumulated rainfall (mm), is obtained from Bangladesh Meteorological Department (1989 to 2014) and Ogimet (1999 to 2018). September is a neutral and transitional month from monsoon to winter to evaluate drought scenarios. Additionally, in Matlab R2018b, Nonlinear Autoregressive with external input (NARX) is tested with several optimization techniques: Levenberg-Marquardt (LM), Bayesian Regularization (BR), and Scaled-Conjugate Gradient. ANN models show that Chittagong has more rainfall than Dhaka supporting climatological statistics. Specifically, forecasts for Dhaka are 25%, 21%, and 22%, and Chittagong 31%, 30%, and 8%, respectively, using LM, BR, and SCG. The iterations for Chittagong 12, 201, and 10 and Dhaka are 5, 12, and 47, respectively, by LM, BR, and SCG. The results suggest rainfall probabilities in September about 20 to 30% of annual events. The study, particularly for Chittagong in ANN, refers to computational resources and time that are significant to test sensitivities before building a meteorological disaster management tool.

Keywords: Agriculture, Artificial intelligence, Bangladesh, Bayesian Regularization, Computation, Conjugate gradient, Climatology, Development, Economy, Levenberg-Marquardt, Meteorological Droughts, Metropolitans, NARX, Neural Network, Optimization, Prediction, Rainfall, Sustainability.