Fires are a leading cause of accidents globally, often resulting in significant destruction, which can be challenging to quantify. It can happen anywhere and under any circumstances, with repercussions that transcend geographical and situational boundaries, causing widespread devastation across the social and economic domains. In this study, we introduced a fire detection model based on a deep learning methodology. The VGG16, VGG19, Inception, and Xception models are widely recognized as standards for image categorization challenges. In our method, we introduced a deep CNN model for fire detection, which was evaluated using a custom dataset. Our model utilized a multipath architecture that incorporated convolutional layers with a combination of small and large filters. This design enables the model to learn local and global features effectively, thereby improving its ability to extract comprehensive features. Varying size filters identify intricate patterns, textures, and structures, leading to the generation of stronger and more expressive representations. We utilized image data gathered from satellite and CCTV surveillance systems deployed in various environments, including forests, agricultural land, indoor and outdoor settings, and chemical plants for fire prevention. The experimental results demonstrate that our proposed model outperforms state-of-the-art methods (VGG16, VGG19, Inception, and Xception) in terms of Accuracy, Precision, Recall, F-score, and ROC-AUC score.