Transparent insights into AI: analyzing CNN architecture through LIME-based interpretability for land cover classification

Abstract Deep learning models demonstrate significant efficiency in remote sensing; however, their transparency presents challenges for operational monitoring applications. This research examines the efficiency and interpretability of various convolutional neural network (CNN) architectures for detecting wind turbines using SPOT-6/7 satellite imagery. Nine CNN architectures were evaluated using a standardized protocol, with MobileNetV2 demonstrating the highest overall performance. The study utilizes Local Interpretable Model-agnostic Explanations (LIME) to produce superpixel-based Saliency heatmaps, thereby advancing beyond simple accuracy assessment and elucidating the spatial evidence that informs MobileNetV2 predictions. This analysis consistently emphasizes turbine morphology, including blades, tower, and shadows, while also revealing systematic failure modes in complex scenarios. The reliability metrics, such as the Brier score and Expected Calibration Error (ECE), indicate that the model’s confidence is consistent with its empirical accuracy. These findings establish a clear and reproducible baseline for explainable wind-turbine detection in very-high-resolution satellite imagery, thereby facilitating the advancement of reliable geospatial AI systems.