Decoding the “Class Attraction” Effect in Satellite Imagery: A Quantitative Explainable AI Approach

Abstract

While deep learning architectures have achieved remarkable success in Land Use and Land Cover (LULC) classification from high-resolution satellite imagery, their inherent “black-box” nature limits their reliability in critical Earth observation applications. Recent literature has identified the “class attraction” effect-a phenomenon where models consistently misclassify spectrally overlapping classes, such as Agriculture and Barren Land. However, traditional evaluation metrics, such as the mean Intersection over Union (mIoU), fail to elucidate the spatial reasoning behind these decision failures. In this paper, we introduce a quantitative Explainable AI (XAI) framework to decode this phenomenon. Using Gradient-weighted Class Activation Mapping (Grad-CAM), we extract the spatial attention of four distinct Convolutional Neural Network (CNN) architectures: U-Net, SegNet, PSPNet, and DeepLabV3+, evaluated on the DeepGlobe dataset. To transcend subjective visual inspection, we propose two novel spatial explainability metrics: Attention-Intersection over Union (A-IoU) and the Energy Ratio. Our quantitative results demonstrate that standard encoderdecoder models like U-Net exhibit severe attention scattering, capturing only 45.0% of their activation energy within the target class boundaries (A-IoU of 0.25). In contrast, DeepLabV3+ demonstrates highly focused spatial reasoning, directing 82.0% of its energy to the correct topological features (A-IoU of 0.68). These findings prove mathematically that multi-scale contextual aggregation, such as Atrous Spatial Pyramid Pooling (ASPP), is essential not only for accuracy but for mitigating the class attraction effect by anchoring the model’s focus on global geometric consistency rather than deceptive local spectral intensities.