A Comparative Study of Spatio-Temporal Segmentation Performance: AWS g4dn.xlarge vs. Google Colab T4 GPU

In this study, we investigated the effect of training image Gaussian noise augmentation on the performance of semantic segmentation task in satellite images and bench-marked the results on two cloud computing platforms: Google Colab (T4-GPU) and AWS (g4dn.xlarge). Using a dataset of 100 each of training, validation, and test patches, we trained models for 15 epochs with a batch size of 2 (with and without Gaussian noise). The results showed notable performance differences between platforms. On the T4 GPU, the model achieved 72.96% accuracy and 31.00% IoU (Intersection over Union) without noise, slightly improving to 74.54% accuracy and 31.56% IoU with noise. In contrast, AWS g4dn.xlarge had better results, with 77.80% accuracy and 38.02% IoU without noise, and 79.15% accuracy and 38.51% IoU with noise. The significant variation, especially in IoU, suggests that hardware differences affect model performance. These findings highlight the importance of choosing appropriate computational resources in deep learning experiments. Future work will explore the relationship between hardware and training efficiency and optimize noise-based augmentation strategies.