International Journal of Science, Engineering and TechnologyAuthors: Gowri Mysuru, Gowri C V, Inchara K Shekhara, Kavana N, Mrs. Lavanya S
Abstract: Forest fire incidents cause severe damage to natural ecosystems, wildlife habitats, and human life, making rapid identification a quibbling requisite for telling excuse and disaster response. Conventional fire monitoring formulation such as enchiridion surveillance, satellite observation, and sensor-based method rarely face inquiring affiliated to delayed response, high operational cost, and limited coverage. To surmount these terminal point, this line proposes an intelligent image-based system for detection of woodland fire using deep learning techniques. The system categorizes forest images into three distinct classes—Fire, Smoke, and Normal—to support early-stage fire recognition. An ensemble of advanced architectures, namely ConvNeXt-Tiny, EfficientNetV2, and Swin Transformer, is exploited to capture fine-grained visual features as well as broader contextual information. A diverse dataset of forest appearance obtained from publicly gettable sources is utilized, along with extensive preprocessing and data augmentation to enhance model strength low-level varied biological science conditions. Input images are resized and normalized before being processed by the trained models, and final predictions are determined using probability-based decision fusion. Experimental evaluation shows that the proposed approach achieves an overall accuracy of 98% on the test dataset, with consistently high precision and recall across all categories. The resolution establish that the system can reliably identify fire and smoke scenarios while reducing false detections, constituent it desirable for real-time forest monitoring and early warning applications.
