Comparative analysis of CNN architectures for satellite-based forest fire detection: A mobile-friendly approach using Sentinel-2 imagery

Abstract

This study evaluates the performance of nine convolutional neural network (CNN) architectures for fire detection using Sentinel-2 satellite imagery from Mount Kilimanjaro National Park. It aims to identify the most effective models by balancing detection accuracy, computational efficiency, and deployment feasibility, especially in resource-constrained environments. The study employed a comparative analysis of traditional (AlexNet, VGG16, VGG19), advanced (ResNet-50, ResNet-101, Inception-v3), and mobile-friendly architectures (MobileNetV2, MobileNetV3, EfficientNet-B2). Despite a limited base set of 60 Sentinel-2 images, we derived 2940 image patches and applied augmentation to support robust model comparison. Performance was evaluated using metrics such as accuracy, precision, recall, and F1-score, while computational efficiency was assessed using FLOPs, inference time, and memory usage. Statistical validation using the Mann-Whitney U test ensured the reliability of the results. MobileNetV2 emerged as the optimal architecture for resource-constrained environments, achieving near-perfect performance metrics (precision, recall, and F1-score of 0.99) with minimal computational requirements (300M FLOPs, 12ms inference time). ResNet-101 demonstrated the highest accuracy (99 %) among advanced models but required substantial computational resources. The results highlight the importance of leveraging multi-spectral data, particularly Sentinel-2's short-wave infrared bands, for accurate fire detection. Statistical validation confirmed significant performance differences among models, with MobileNetV2 and ResNet-101 outperforming alternatives in their respective categories. While the evaluation focused on one ecological region and year, future work will extend this analysis across time and geography for broader generalization. This study bridges the gap between computational advancements and practical deployment needs by providing actionable insights into CNN model selection for real-time fire detection systems. It uniquely combines Sentinel-2's multi-spectral capabilities with advanced machine learning models, offering a scalable framework for addressing environmental challenges in resource-limited settings. The findings contribute to sustainable fire management practices and open new avenues for deploying CNNs in environmental monitoring.