Empirical analytics of baseline and enhanced CNN architectures with frequency and RGB features using Bayesian Hyperparameter Optimization for wildfire prediction

­Wildfires have caused long-term economic, ecological, and biological damage, highlighting the need for accurate prediction systems to protect forest wildlife and valuable non-timber resources such as medicinal plants, aromatic products, food, fodder, and fuelwood. This study proposes wildfire prediction models using frequency-domain analytics and Bayesian Optimization (BSO) in designing Convolutional Neural Network (CNN)-based deep learning models, including LeNet-5, AlexNet, and VGG16, applied to the DeepFire dataset. The models are trained and tested on both RGB images and Fast Fourier Transform (FFT)-based frequency-domain representations of fire and non-fire images. BSO, integrated with the Tree-structured Parzen Estimator (TPE), optimizes model parameters to effectively extract fire-related features. Model performance is evaluated using Accuracy and AUC-ROC metrics. Results indicate that BSO-based frequency-aware modified LeNet-5 and AlexNet achieve accuracies of 97% and 96%, respectively. Additionally, RGB-based BSO enhances performance, with modified LeNet-5 and baseline VGG16 reaching up to 98% accuracy. Overall, findings demonstrate that frequency-domain features combined with BSO significantly improve wildfire prediction, supporting ecological and biological conservation efforts.