Forest fire detection is essential for an efficient and rapid response to environmental protection. Existing methods for wildfire detection often rely on traditional image processing techniques or shallow learning models, which struggle with real-time performance and accuracy, particularly in remote forest environments with limited processing resources. This study addresses the performance gap in forest fire detection methodologies by comparing multiple convolutional neural network architectures. Our objective was to identify the most accurate and efficient model for real-time wildfire detection from RGB image data. Deep learning models such as CNN, ResNet50, EfficientNetB0, EfficientNetB3, SqueezeNet, ShuffleNet, and MobileNetV2 are evaluated on a benchmark Wildfire dataset. Further, we proposed a lightweight MobileNetV2 approach fine-tuned on the ImageNet dataset. The proposed fine-tuned MobileNetV2 approach outperformed with an accuracy of 0.93. These findings significantly affect the development of low-power, high-performing fire detection systems that could be deployed in remote forest environments with limited processing resources. The research done can be used for existing automated wildfire monitoring technologies, enabling better forest management and conservation strategies.