Adaptive Hierarchical Multi-Headed Convolutional Neural Network with modified Convolutional Block Attention for aerial forest fire detection

Effective detection and classification of forest fire imagery are critical for timely and efficient wildfire management. Convolutional Neural Networks (CNNs) have demonstrated potential in this domain but encounter limitations when addressing varying scales, resolutions, and complex spatial dependencies inherent in wildfire datasets. Building upon our prior work on the Unmanned Aerial Vehicle-based Forest Fire Database (UAVs-FFDB) and the multi-headed CNN (MHCNN), this study introduces a novel architecture, namely, the Adaptive Hierarchical Multi-Headed Convolutional Neural Network with Modified Convolutional Block Attention Module (AHMHCNN-mCBAM). This enhanced framework addresses prior challenges by integrating adaptive pooling, concatenated convolutions for multi-scale feature extraction, and an improved attention mechanism incorporating shared fully connected layers, Glorot initialization, rectified linear units (ReLU), layer normalization, and attention-gating. AHMHCNN-mCBAM incorporates Gated Recurrent Unit (GRU) and Bidirectional Long Short-Term Memory (BiLSTM) networks for temporal context modeling to further refine classification accuracy. Experiments conducted on the UAVs-FFDB dataset achieved outstanding results, including 100% accuracy, a 100% Cohen’s kappa coefficient (cKappa), and compact model parameter sizes of 1.49 million (M), 0.25 M, and 0.12 M. On the Fire Luminosity Airborne-based Machine Learning Evaluation (FLAME) dataset, the model attained accuracy rates of 99.83%, 99.10%, and 99.32%, with corresponding cKappa values of 99.66%, 98.20%, and 98.65%. Compared to the baseline hierarchical MHCNN with CBAM (HMHCNN-CBAM), AHMHCNN-mCBAM demonstrated significant performance gains, including a 6.80% and 6.59% increase in accuracy, a 9.26% and 14.11% improvement in cKappa, and a 13.87% and 13.76% reduction in parameter size on the UAVs-FFDB and FLAME datasets, respectively. Additionally, AHMHCNN-mCBAM outperformed HMHCNN-CBAM in recall (25% improvement), precision (21.95%), F1-score (14.94%), and false discovery rate (FDR) reduction (25.01%), while achieving a 100% reduction in error rate (EWR). Leveraging Explainable Artificial Intelligence (XAI) techniques, the model provides interpretable insights into decision-making processes.