Lithological mapping with multispectral remote sensing remains challenging when diagnostic spectral information is limited and reliable labeled samples are scarce. This problem is particularly relevant when convolutional neural networks (CNNs) are applied to lithological classification, because limited spectral dimensionality and scarce training samples may hinder the learning of discriminative spatial–spectral features. In this study, we developed a limited-sample lithological mapping framework for the Shibaocheng area of Subei County, Gansu Province, China, using band-integrated ASTER and Sentinel-2A multispectral imagery. ASTER shortwave infrared (SWIR) bands were co-registered and resampled to Sentinel-2A imagery, and then integrated with Sentinel-2A visible and near-infrared (VNIR) and red-edge bands to construct a complementary multispectral dataset. A compact spectrally enhanced multi-scale CNN was designed, incorporating a residual spectral feature enhancement module for inter-band representation learning and a parallel multi-scale hybrid convolution module for capturing spatial–spectral features. Eight lithological units were classified under limited-label conditions using 8158 training samples and 3497 spatially independent validation samples. Experimental results show that the band-integrated ASTER–Sentinel-2A dataset improved classification performance compared with single-sensor inputs. Using the proposed model, the band-integrated dataset achieved an overall accuracy (OA) of 94.12%, average accuracy (AA) of 94.04%, and Kappa coefficient of 0.932, compared with OA values of 93.14% and 92.40% obtained using ASTER and Sentinel-2A alone, respectively. The positive effect of band-level integration was also observed for spectral angle mapper (SAM), support vector machine (SVM), and 3D-CNN, whose OA values increased to 54.33%, 86.12%, and 92.29%, respectively. The proposed CNN achieved the highest OA among the evaluated methods, outperforming SAM, SVM, and the conventional 3D-CNN. In addition, t-SNE visualization indicated that incorporating spatial texture features produced more compact and better-separated lithological clusters than using spectral features alone. Ablation experiments further demonstrated that the proposed spectral feature enhancement and multi-scale hybrid convolution modules each contributed to improving lithological classification performance. These results demonstrate that integrating freely available multispectral data with a lightweight spectral–spatial CNN provides a practical and cost-effective solution for lithological mapping in bedrock-exposed arid to semi-arid regions, especially where hyperspectral imagery and dense field samples are unavailable.
Pei et al. (Fri,) studied this question.