CNN-MLP framework for forest burned areas prediction using PSO-WOA algorithm

Abstract

Accurate prediction of the burned area from forest fires is essential for effective wildfire risk management and mitigation; however, this task remains challenging due to the highly nonlinear relationships and extreme skewness inherent in fire-weather data. This study proposes an optimized hybrid deep learning framework that integrates a Convolutional Neural Network and a Multilayer Perceptron (CNN–MLP) with metaheuristic optimization to enhance burned-area regression accuracy. A wrapper-based Binary Firefly Algorithm (BFA) is employed to identify the most informative subset of meteorological, fire-weather, and spatial features, reducing redundancy and improving generalization. To further optimize model performance, a hybrid Particle Swarm Optimization–Whale Optimization Algorithm (PSO–WOA) is used to tune the hyperparameters of the CNN–MLP architecture automatically. Experiments conducted on the UCI Forest Fires dataset demonstrate that the proposed PSO–WOA–CNN–MLP model significantly outperforms optimized baseline deep learning models, including CNN, MLP, LSTM, and GRU. The proposed model achieves very low prediction errors (Mean Squared Error (MSE) = 0.0093, Mean Absolute Error (MAE = 0.0767, Median Absolute Error (MedAE = 0.0616, Mean Absolute Percentage Error (MAPE = 0.0066) and an exceptional coefficient of determination (R² = 99.89%), indicating near-perfect agreement between predicted and observed burned areas.