MESNET: integrating lightweight CNNs and real-time carbon tracking for sustainable image classification

Abstract The rapid growth of artificial intelligence has raised increasing concern about its environmental impact, particularly the energy consumed during model training and inference. This study presents MESNET, a lightweight ensemble framework that integrates three convolutional neural network architectures (MobileNetV3-Large, EfficientNetV2-S, and ShuffleNetV2) with real-time carbon emission tracking for sustainable image classification. Experiments were conducted on the CIFAR-10 dataset using a standardized cloud-based environment (NVIDIA Tesla T4 GPU, PyTorch, CodeCarbon). The framework evaluates models not only by accuracy but also by computational cost (GFLOPs), parameter count, energy consumption, and CO₂ emissions. Among individual models, MobileNetV3-Large achieved a strong balance between performance and efficiency, while ShuffleNetV2 demonstrated the lowest FLOPs and emissions. By combining their complementary strengths, the proposed MESNET ensemble achieved 96.98% accuracy, 96.98% F1-score, and produced only 6.863 × 10⁻ 3 kg of CO₂ during training. Energy-normalized metrics further confirmed MESNET as the most efficient model, delivering 6,633 F1-score points per kWh. These results demonstrate that high classification accuracy and environmental sustainability are not mutually exclusive, and that MESNET provides a reproducible and scalable pathway toward Green AI.