Bayesian Optimization Based Deep Learning Models for Detection of Forest Fires

Abstract

Fires that may break out in forest areas can be a great threat to people and biodiversity. This derives the need for effective fire detection systems that analyze images captured by watchtowers. In this study we employ a computer vision-based deep learning approach to classify images for early detection of forest fires. A benchmark dataset known as Kaggle's forest fire dataset was used for training and testing purposes. Deep learning CNN models such as AlexNet, ResNet18, DenseNet121 and EfficientNetb0 were trained and evaluated as binary classifiers. To optimize model performance hyperparameter tuning was performed using Bayesian optimization, which efficiently explores the search space by modelling the objective function. Additionally, k-fold cross-validation was applied to ensure stable estimates and validation metrics of different folds were averaged for each of the four models. After selecting the best weights for each model, test data was used to obtain the respective performance metrics. The paper also provides results for majority voting and probability averaging which are ensemble methods. Finally, the findings are compared against those of three other studies. © 2025 IEEE.