<![CDATA[This study explores the implementation of a Convolutional Neural Network (CNN) for automated forest fire identification using digital image processing. Utilizing the USTC 'Forest Fire' dataset, the research framework involved systematic data preprocessing, including a 70:30 training-validation split and the application of image augmentation techniques to enhance model robustness. The proposed architecture features a sequential design with dual convolution and pooling layers, integrated with ReLU and Sigmoid activations. Although initial training over seven epochs yielded a deceptive validation accuracy of 99%, granular performance analysis exposed critical limitations. Evaluation via a Confusion Matrix revealed that while the model excelled at identifying 'non-fire' scenarios, it struggled significantly with actual fire detection, failing to recognize 301 out of 331 fire instances. These results highlight a severe class imbalance issue, suggesting that standard accuracy metrics are insufficient for this application and emphasizing the need for more balanced sampling or advanced architectural adjustments in future fire detection systems.]]>
Copyrights © 2026
