DEEP LEARNING-DRIVEN WILDFIRE DETECTION: A HYBRID FRAMEWORK INTEGRATING MULTI-SOURCE DATA AND ENSEMBLE LEARNING

Abstract

This paper presents a state-of-the-art framework for detection and prediction of forest fires with the use of advanced ensemble learning techniques in combination with CNNs. Traditional systems of forest-fire detection suffer from problems of poor accuracy, slow speeds, and scalability issues. On the contrary, this paper uses a hybrid approach which integrates DL and DIP specifically for handling multi-source datasets comprising satellite imagery, drone videos, and environmental variables like temperature, wind velocity, and humidity. The data preprocessing techniques that will be included for improving feature extraction effectiveness will involve color space conversion, image enhancement, and spatial segmentation. The CNN architecture had a number of convolutional and pooling layers that were set up to extract high-dimensional features pertinent to fire detection. Transfer learning using pre-trained models significantly improved performance even with limited labeled data. This was complemented by ensemble learning where the outputs of CNN models were combined with probabilistic models such as Random Forests and Gradient Boosting for enhancement of classification robustness and avoidance of overfitting. The system was trained on different datasets collected from the Kaggle repositories, satellite feeds, and field observations of the Margalla Hills. A k-fold cross-validation technique ensured the generalization of the model, and it achieved 92% detection accuracy with a precision of 89% and an F1-score of 90%. Complex scenarios such as smoke fog differentiation and small-scale fire hotspots in dense vegetation were successfully addressed. Experimental results show that the model has better adaptability in a variety of environmental conditions, and its real-world evaluation proves its validity. Predictive capabilities further enabled identification of fire-prone zones before ignition while triggering its utility in early intervention systems.