Abstract Effective mitigation and response to disasters require fast and accurate classification and localisation of natural disasters. We herein propose EcoDisasterLocNet, a hybrid architecture of Grad-CAM + + and DenseNet-201 that integrates fine-grained and interpretable disaster localisation with accurate spatial classification. The architecture is learned on a balanced and augmented 10,537-image dataset spread over four classes: earthquake, wildfire, flood, and cyclone. 75% of the images were utilised for training while 15% and 10% of the images respectively tested and validated the architecture. Optimised DenseNet-201 produced the best results where it obtained 99.90% training accuracy, 94.24% validation accuracy, and 95.16% testing accuracy along with 95.26%/95.16%/95.19% precision/recall/F1-score. This hybrid ensemble (ERI-2025) was initially evaluated alongside standalone CNN models. Grad-CAM and Grad-CAM + + visualisations were used to emphasise disaster-related Grad-CAM and Grad-CAM + + visualisations were used to emphasise disaster-related regions (e.g., fissures, wildfires, flood regions, cyclone eyes) for localisation. Grad-CAM + + enhanced the IoU from 0.27 to 0.51 to 0.39–0.51 while maintaining a Dice coefficient of 2.00. The superiority of DenseNet-201 in both classification and localisation tasks was confirmed by comparative evaluation against other CNN architectures The hybrid design of the proposed framework is adaptable to multimodal and spatiotemporal datasets in future research, and it provides a scalable, interpretable, and high-precision solution for real-time disaster monitoring.
Raju et al. (Wed,) studied this question.