Accurate wetland classification demands high spectral detail and large-area coverage – capabilities that neither satellite multispectral data (limited spectral resolution) nor unmanned aerial vehicle (UAV) hyperspectral data (limited spatial coverage) can provide simultaneously. To alleviate this deficiency, this study developed a hyperspectral data simulation framework that fused UAV hyperspectral data with satellite multispectral data to enable fine-scale land cover classification of the Qilihai Wetland. UAV hyperspectral data were acquired over multiple representative land-cover patches and integrated with Ziyuan-1 02E (ZY1F) multispectral data. A deep learning model was used to learn local nonlinear mappings from multispectral to hyperspectral space, thereby enabling accurate spectral reconstruction and spatial extrapolation in areas lacking hyperspectral coverage. This method ultimately generated quasi-hyperspectral data covering the entire wetland, featuring a spatial resolution of 2.5 m and 300 spectral bands. Quasi-hyperspectral data were validated using evaluation metrics at both band and pixel levels, and the simulation results showed low spectral angle mapper values (generally ≤0.1724). Subsequently, the quasi-hyperspectral data were used for Qilihai Wetland land cover classification. The results yielded an overall accuracy (OA) of 90.59%, a Kappa coefficient of 0.8632, and an average accuracy (AA) of 89.41%, outperforming the use of multispectral data alone by 9.52%, 0.1376, and 8.15%, respectively. The proposed framework mitigates the limited spectral resolution of multispectral imagery while extending hyperspectral information beyond the UAV coverage, providing a scalable multi-source fusion solution that can be adapted through site-specific calibration to support wetland monitoring and management.
Shen et al. (Sun,) studied this question.