Abstract:
Hyperspectral remote sensing technology has garnered significant attention in the field of wetland remote sensing classification. Selecting appropriate characteristic bands is of paramount importance for wetland classification. In this study, the Jiangxi Poyang Lake Nanji Wetland National Nature Reserve was chosen as the study area due to its ecological diversity and representative wetland characteristics. Comprehensive analysis of hyperspectral characteristics for water bodies, vegetation, and other land cover types within the wetland was conducted using hyperspectral data from the Zhuhai-1 satellite. The characteristic bands for these three major terrestrial features were meticulously screened using the error range threshold method, ensuring only the most relevant and distinct bands were selected. The applicability of the selected characteristic bands was rigorously assessed through the utilization of Mahalanobis distance, a statistical measure that quantifies the spectral differences between features. Subsequently, the random forest classification algorithm was applied to the Nanji Wetland based on the optimal combination of different characteristic bands. The research findings indicated that water bodies retained key wavelengths such as 500 nm and 596 nm through feature optimization, while vegetation spectral characteristics shifted towards the visible light region, effectively capturing the "green peak" and "red edge" features. For other terrestrial features, after spectral transformation, the bands were concentrated at 531 nm, 560 nm, 596 nm, etc. For the same type of land features, the Mahalanobis distance values were comparatively small on the feature bands obtained after three kinds of feature transformations. The combination of feature bands significantly strengthened the spectral differences among terrestrial features, reducing misclassification and omission phenomena. Notably, the overall classification accuracy of the combination of original data, continuum removal, and first-order derivative transform feature bands reached 92.0%, with the Kappa coefficient as high as 0.9046, which were 42.41% and 52.83% higher than in terms of overall classification accuracy and Kappa coefficient using the characteristic bands from original data alone, respectively. The research results can not only provide a theoretical basis for the selection of feature bands in wetland remote sensing classification, but also provide technical references for the identification and monitoring of inland lake wetlands.