<![CDATA[Indonesia is one of the countries with a high incidence of forest and land fires (karhutla), especially during the dry season, thus requiring a fast and efficient early detection system. This study aims to compare the performance of two popular deep learning architectures, namely MobileNetV3 (Large and Small variants) and ResNet50, in forest fire image classification tasks using a transfer learning-based approach. This study emphasizes the comparison between accuracy and computational efficiency in a CPU-only environment, which represents real-world conditions of use in the field without GPU support. The dataset used is a combination of local field images from the Puncak area, Bogor, and a curated public forest fire dataset to ensure the model's generalization ability to diverse geographical conditions. The results of the experiment show that ResNet50 provides the highest accuracy with a training accuracy value of 0.677 and a validation accuracy of 0.647, but requires longer training and inference times. Meanwhile, MobileNetV3-Large and MobileNetV3-Small showed better computational efficiency, with only slightly lower accuracy (0.635 and 0.61) and high training stability. These findings confirm that lightweight models such as MobileNetV3 strike an optimal balance between accuracy, speed, and resource consumption, making them an ideal solution for implementing edge computing-based early detection systems. Overall, this research contributes by providing an empirical comparative analysis that can serve as a reference for selecting deep learning architectures for efficient and adaptive forest fire detection systems that are constrained by hardware limitations.]]>
Copyrights © 2025
