Publication

Abstract : Forest fires are a critical environmental concern, causing extensive damage to ecosystems, biodiversity, and human livelihoods, necessitating accurate detection systems. This study investigates the performance of CNN, specifically ResNet50, for forest fire classification under two scenarios: (1) using the same dataset for validation and testing, and (2) employing distinct datasets for training, validation, and testing. To elucidate model behaviour, an analogy to a student’s learning process is used, where training represents learning, validation reflects practice, and testing mimics exams. In the first scenario, akin to a leaked exam, ResNet50 achieved high accuracies (96.88% for training, validation, and testing); however, its classification metrics were suboptimal, with precision, recall, and F1 scores of 0.46 for both fire and no-fire classes, indicating overfitting and memorization. In the second scenario, distinct dataset partitioning addressed overfitting, resulting in consistent accuracies (96.88% for training and testing, 96.43% for validation) and robust classification metrics (fire class: precision 0.97, recall 0.96, F1 score 0.97). These findings highlight the importance of appropriate dataset partitioning in ensuring model generalization and reliability, demonstrating ResNet50’s potential as a robust tool for forest fire classification when guided by well-structured experimental designs.