Toward Explainable AI in Satellite Imagery: A ResNet-50-Based Study on EuroSAT Classification

Deep learning has substantially advanced satellite image classification, yet the opaque nature of neural networks continues to impede interpretability in remote sensing applications. While convolutional neural networks, such as ResNet-50, consistently demonstrate high classification accuracy, understanding their internal decision-making processes, particularly across heterogeneous land use categories, remains a critical challenge. This paper examines the interpretability of deep learning-based classification models using two prominent Explainable AI techniques: Gradient-weighted Class Activation Mapping and Local Interpretable Model-agnostic Explanations. Leveraging a pre-trained ResNet-50 model on the EuroSAT dataset, we analyzed twenty representative satellite image samples, comprising both correctly and incorrectly classified instances across ten land use classes. Grad-CAM generated class-discriminative spatial attention maps that identified salient regions influencing the model’s predictions, while LIME provided super pixel-based visual explanations reflecting localized feature contributions. Comparative analysis of both techniques revealed consistent patterns of model attention and exposed spectral and spatial ambiguities that underlie misclassification. These insights not only enhance the interpretability of CNN-based remote sensing models but also inform the development of more transparent and reliable Earth observation systems.