Transparent Insights into AI: Analyzing CNN Architecture through LIME-Based Interpretability for Land Cover Classification

The realization that complex deep learning models may make morally significant decisions has led to a growing interest in Explainable Artificial Intelligence (XAI), whose primary concern is understanding why it made particular predictions or recommendations. This paper investigates the effectiveness of different Convolutional Neural Network (CNN) architectures that are employed on satellite images from the Airbus SPOT6 and SPOT7 Datasets. The evaluated designs are MobileNetV2, Alex Net, ResNet50, VGG16, DenseNet, Inception-ResNet v2, InceptionV3, XceptionNet, and EfficientNet. MobileNetV2 showed best in other classification parameters such as accuracy of 99.20%, precision rate of 99.39%, recall rate of 99.00 %, F1 score to be at a maximum with 99.16 % and an AUC (Area Under the Curve) to be detected across all categories correctly at 99.96%. The research study uses LIME (Local Interpretable Model-agnostic Explanations) to examine MobileNetV2, a system that uses satellite images to classify wind turbines. LIME creates interpretable models, such as white box models, to estimate complex predictions. This helps identify key factors in classification, making the model more interpretable. The study uses heatmaps and attention maps to identify areas in Airbus SPOT satellite images that impact MobileNet classifications. This enhances trust in the AI system and opens up opportunities for understanding model behaviour.