Hybrid Deep Learning Framework Integrating EfficientNet-B0 and YOLOv11 for Robust Forest Fire Detection and Classification

Wildfires demand quicker and reliable detection for effective intervention. This work presents a compact deep learning system that unifies global scene classification and local fire/smoke detection using EfficientNet-B0 and YOLOv11, connected through a fusion layer. On the benchmark Kaggle (4,823 images) and Roboflow (2,382 images) datasets, our approach reached 98.6% classification accuracy, F1-scores exceeding 0.98, and a detection precision of 64% (mAP 0.64). These results show a clear improvement over prior models, with the fusion logic sharply reducing false positives in ambiguous images. Ablation studies confirm robust generalizable results. Due to its efficiency, the model is ready for real time deployment with UAVs and IoT devices in resource constrained environments, marking progress in practical wildfire monitoring.