Multi-Scale Spectral-Spatial Attention and Conditional Color-Space GAN Synthesis for Land Cover Classification

Résumé fourni par la source

Satellite and aerial imagery are required to classify land cover with high accuracy to provide climate resilience, urban planning, and environmental monitoring, but spectral overlap among classes, spatial-scale variability, and limited labeled data have made this difficult. To solve these problems, this paper will introduce a hierarchical spectral-spatial attention network with conditional generative data augmentation. The model also includes multi-scale spectral attention of channel-wise re-calibration of features and scale-sensitive spatial attention to distinguish receptive field contextual dependencies. To alleviate data scarcity, a Multi-modal Color-Space GAN (MMCS-GAN) is proposed, which produces paired RGB and HSV images with explicit color consistency regularization to enhance semantic realism and balance between classes. Experiments on a 7-class land cover dataset indicate that the proposed spectral-spatial framework has an accuracy of 95.62, good Macro F1 (0.9564), and weighted F1 Scores. RGBHSV consistency offers a significant addition to RGB-only augmentation, and hierarchical attention adds about 2.3 to the performance. The model also exhibits excellent robustness, where the accuracy decreases when there is noise and brightness perturbation, and the calibration is also enhanced as indicated by a smaller log loss of 0.1801. The excellence of the proposed method over baseline methods is further confirmed by statistical tests, such as the calibration curves and analysis of significance. The framework offers a powerful, generalizable method of high-resolution land cover mapping that uses fewer manual annotations.