This study addresses ambiguity in spatial extent and boundaries in satellite image classification to improve the accuracy of fine-grained object-based Level 2 land cover classification. Unlike conventional data augmentation, we propose a novel Style-Adaptive U-Net that incorporates the visual characteristics of landscape paintings into classification learning. Specifically, we developed a lightweight CNN-based Art Encoder coupled with an Enhanced Style Feature Fusion (ESFF) module to inject artistic features into the network’s feature representation. Based on visual features extracted from works by Egon Schiele, Van Gogh, Claude Monet, and Elyse Dodge, the model utilizes painting styles with distinct boundaries or strong textures to explicitly enhance the boundary recognition capability of objects. Experimental results demonstrate the efficiency and superiority of the proposed model. It achieves a peak Dice score of 0.7631, outperforming the baseline U-Net’s 0.6512, and maintains a manageable processing load with only a 19% increase in parameters. Our comparative analysis shows a distinct representational mechanism by demonstrating that styles with explicit structural features (Schiele, Dodge) improve boundary discrimination, whereas styles emphasizing blurred transitions (Monet) yield limited functional gain. This validates our premise that the network actively utilizes artistic features as functional structural guidance rather than mere aesthetic enhancements, offering an efficient paradigm for resolving geographic ambiguity.
Won et al. (Wed,) studied this question.