Abstract:

Accurate identification of wetland vegetation is critical for monitoring and protecting wetland ecosystems.However,traditional multispectral remote sensing struggles to distinguish wetland plants with similar spectral signatures because of its broad spectral bands and cannot capture spectral changes across different phenological stages.Hyperspectral data,with its continuous and fine spectral resolution,provide a solution for detailed plant classification.Nevertheless,the variation patterns of spectral characteristics of different plants during phenological changes remain unclear,limiting the accuracy of cross- temporal classification.To investigate the difference in spectral characteristic of wetland plants under phenological changes and their impact on cross-temporal classification accuracy,the Sanhekou Wetland,where the Yellow River,Wei River,and Beiluo River converge,was selected as the study area.Field hyperspectral observations were conducted during vigorous growth stage (June) and senescence stage (September) in 2022.Using the envelope removal,“three-edge”parameter extraction,vegetation index construction,and PCA-CNN feature fusion,spectral differentiation patterns were analyzed and cross-phase classification models were established.The results indicated that:1) Plants with reflectance less than 45% at 750 nm and a red-edge reflectance increment less than 0.3 per 10nm were identified as woody plants,while those with the opposite characteristics were identified as herbaceous plants.The red-edge chlorophyll index (RECI) and red-edge normalized difference vegetation index (RENDVI) exhibited higher sensitivity for distinguishing plant types than conventional vegetation indices.2) The PCA-CNN feature fusion model achieved a classification accuracy of 66.24%,significantly improving classification performance compared with single-feature models.3) The model trained using June data showed a significant drop in accuracy to 41.1% when tested with September data.This decline was mainly attributed to the asymmetric spectral responses of different plants to phenological stages,which weakened the cross-temporal generalization capability of the model.These results reveal that wetland vegetation exhibits distinct spectral differentiation patterns during phenological changes.The PCA-CNN feature fusion model effectively improves fine-scale wetland vegetation classification performance,and phenological changes are an important factor affecting cross-temporal classification accuracy.This study provides technical support for fine-scale identification of regional wetland plants and large-scale hyperspectral remote sensing classification.

Key words: wetland vegetation, hyperspectral remote sensing, phenological change, feature fusion