Bangladesh’s annual battle with water is entering a new technological era. A research team from Khulna University of Engineering and Technology has developed a hybrid rainfall-modeling framework that combines wavelet signal processing, seasonal decomposition, artificial neural networks, Bayesian optimization, and several nature-inspired algorithms. Tested on rainfall records from Sylhet and Chittagong, two regions repeatedly exposed to intense precipitation and flooding, the approach is designed to extract useful patterns from data that are notoriously nonlinear, noisy, and difficult to predict. The study, published in Neural Computing and Applications, reports that the Harris Hawks Optimization algorithm produced particularly strong results, suggesting that algorithms inspired by animal behavior may help improve weather intelligence in climate-vulnerable regions.
The challenge begins with the character of rainfall itself. Unlike a steadily changing physical variable, daily precipitation often arrives as a mixture of long-term climate trends, seasonal cycles, short-lived atmospheric disturbances, and apparently random extremes. A single day may be dry, followed by an intense cloudburst, while the broader monsoon rhythm remains visible across months or years. This combination makes rainfall difficult for conventional statistical models and ordinary neural networks to represent. In Sylhet, heavy rainfall interacts with low-lying haor wetlands and upstream water flows to intensify flood risks. Chittagong, meanwhile, faces the additional complications of steep terrain, rapid runoff, urban expansion, and coastal weather systems. More accurate rainfall modeling could therefore support flood warnings, reservoir planning, agricultural decisions, drainage management, and emergency preparation.
The framework proposed by Shuvendu Pal Shuvo and Md. Jahir Uddin begins with long-term daily rainfall data collected through Google Earth Engine. Rather than feeding the raw series directly into a predictive model, the researchers first apply a two-stage preprocessing strategy. The first stage uses Seasonal Trend Decomposition, or STD, to separate the observed rainfall signal into broad components such as trend, recurring seasonal behavior, and irregular residual variation. In time-series analysis, this separation is valuable because a neural network can struggle when slow-moving annual cycles and rapid daily fluctuations are mixed together. By isolating the different sources of variation, the model receives a more structured representation of the atmosphere’s behavior.
The second stage applies the Discrete Wavelet Transform, commonly known as DWT. Wavelets are mathematical functions that allow a signal to be examined simultaneously in time and frequency. Fourier analysis can show which frequencies exist in a series, but it is less effective at identifying when a short-lived event occurs. Wavelet analysis is better suited to rainfall because it can distinguish gradual changes from abrupt spikes. In this study, the researchers use the Daubechies 4, or DB4, mother wavelet and decompose the data into approximation and detail coefficients. The approximation component captures the smoother, lower-frequency behavior, while the detail coefficients represent progressively finer fluctuations. This multiscale view is intended to suppress noise without erasing the extreme variations that make rainfall forecasting practically important.
Once the rainfall signal has been decomposed, the resulting components are modeled with an artificial neural network. Neural networks are flexible computational systems made from interconnected processing units that learn relationships between inputs and outputs by adjusting internal weights. For rainfall modeling, those weights determine how strongly different lagged observations and decomposed signal components influence the predicted value. The difficulty is that a network may contain a very large number of possible parameter combinations. Standard gradient-based training can become trapped in a suboptimal solution, particularly when the data are noisy or the error surface is irregular. The researchers therefore use optimization algorithms to search for improved neural-network parameters and model configurations.
The study compares four established or emerging optimization strategies: Particle Swarm Optimization, Genetic Algorithm, Egret Swarm Optimization, and Harris Hawks Optimization. Particle Swarm Optimization models the collective movement of a swarm, with candidate solutions adjusting their positions according to their own experience and the best-known position in the group. Genetic Algorithms borrow concepts from biological evolution, using selection, crossover, and mutation to produce new candidate solutions. Egret Swarm Optimization is inspired by the hunting behavior of egrets and is designed to balance exploration of a search space with exploitation of promising solutions. Harris Hawks Optimization imitates the cooperative hunting tactics of Harris’s hawks, shifting between exploratory movements and increasingly focused attacks as the search progresses.
Bayesian Optimization is incorporated as a probabilistic layer within the broader framework. Unlike a brute-force search, which tests a large number of possibilities directly, Bayesian Optimization builds a surrogate model of the relationship between model settings and prediction error. It then uses an acquisition strategy to decide which configuration should be tested next, concentrating computational effort where improvement is most likely. This is especially useful when evaluating a model is expensive because every candidate configuration may require repeated neural-network training. In practical terms, the probabilistic approach attempts to find effective hyperparameters—such as learning settings, network structure, or optimization controls—without examining every possible combination.
The researchers evaluate performance using Mean Squared Error, Root Mean Squared Error, Mean Absolute Error, and the coefficient of determination, known as R-squared. Mean Squared Error penalizes large mistakes strongly, making it sensitive to failures during heavy-rainfall events. Root Mean Squared Error expresses the same type of error in the original measurement scale, while Mean Absolute Error provides a more direct average of prediction deviations. R-squared indicates how much of the observed variation is explained by the model, although a high value alone does not guarantee that a system will perform reliably during rare extremes. According to the reported findings, the hybrid models achieve very low error values for both Sylhet and Chittagong, with Harris Hawks Optimization delivering particularly strong performance.
The study also examines how sensitive the results are to the choice of wavelet. This matters because wavelets are not interchangeable filters: each mother wavelet has a different mathematical shape, degree of smoothness, and ability to represent sudden changes. The analysis indicates that changing the wavelet can significantly affect model errors, demonstrating that signal preprocessing is not a minor technical detail but a central modeling decision. At the same time, the Harris Hawks-based system reportedly remains comparatively robust across different wavelet choices in both locations. That stability is important for researchers attempting to transfer the method to other stations or climatic settings, where the best decomposition may not be known in advance.
To test whether the results extend beyond a single division of the data, the researchers use fivefold cross-validation. In this procedure, the dataset is divided into five portions, with the model trained on four and evaluated on the remaining portion. The process is repeated until each portion has served as a validation set. This helps reveal whether a model has simply memorized the training data, a danger that is especially serious when flexible neural networks are combined with powerful optimizers. The reported cross-validation results support the generalization and robustness of the optimized framework. Nevertheless, rainfall prediction remains a moving target: a model trained on historical observations may face future conditions shaped by urbanization, changing land use, warming oceans, altered circulation patterns, and climate-driven shifts in extreme rainfall.
The significance of the work extends beyond an algorithmic contest between PSO, GA, ESOA, and HHO. It presents a complete chain for transforming chaotic environmental observations into a structured predictive system. Seasonal decomposition identifies recurring organization, wavelets expose behavior across time scales, neural networks learn nonlinear relationships, Bayesian Optimization reduces the cost of tuning, and nature-inspired algorithms search for parameter combinations that conventional training may miss. The architecture is therefore less like a single forecasting model and more like a coordinated scientific instrument for interrogating rainfall data. Its strongest potential may lie in decision-support systems that combine predictions with real-time river levels, soil moisture, radar observations, satellite measurements, and floodplain maps.
The authors emphasize that the framework could be useful for Bangladesh’s flood-prone communities, especially in regions where intense rain can quickly overwhelm natural and engineered drainage systems. More accurate forecasts could give local authorities additional time to issue warnings, organize evacuations, protect infrastructure, and manage agricultural operations. However, the study should be viewed as a modeling advance rather than a complete operational forecasting service. The article does not establish that the system can predict every extreme event, nor does it replace physical meteorological forecasting or dense ground-based observations. Its data are also to be made available only on reasonable request, which may limit independent replication until researchers can access comparable records and implementation details.
Even with those limitations, the research arrives at a moment when climate adaptation increasingly depends on extracting reliable signals from imperfect data. For communities facing repeated floods, the difference between a useful forecast and an unusable one may be measured in hours, hectares, and lives. The combination of STD, DB4 wavelets, neural networks, probabilistic tuning, and Harris Hawks Optimization offers one possible route toward more responsive rainfall intelligence. Its viral appeal is easy to understand: a computational model inspired by soaring predators is being used to study storms over Bangladesh. Its scientific value, however, rests on the deeper principle behind the imagery—that carefully designed hybrid systems may be better equipped than isolated methods to interpret the complicated, multiscale signature of a changing atmosphere.
Subject of Research: Rainfall modeling and forecasting in Sylhet and Chittagong, Bangladesh, using signal decomposition, artificial neural networks, Bayesian Optimization, and nature-inspired metaheuristic algorithms.
Article Title: Nature-inspired metaheuristic and probabilistic conjugate solutions for rainfall modeling with wavelet neural network
Article References: Shuvo, S. P., & Uddin, M. “Nature-inspired metaheuristic and probabilistic conjugate solutions for rainfall modeling with wavelet neural network.” Neural Computing and Applications, 38, Article 701 (2026). https://doi.org/10.1007/s00521-026-12415-2
Image Credits: AI Generated
DOI: 10.1007/s00521-026-12415-2
Keywords: Rainfall forecasting, Bangladesh, Sylhet, Chittagong, seasonal trend decomposition, discrete wavelet transform, DB4 wavelet, artificial neural network, Bayesian Optimization, Harris Hawks Optimization, Egret Swarm Optimization, Particle Swarm Optimization, Genetic Algorithm, flood prediction, metaheuristic optimization
