Assessment of plant diversity index in degraded desert grassland using UAV hyperspectral multimodal data and Encoder-CNN

The biodiversity function of the desert steppe ecosystem faces many challenges under the pressure of climate change and human activities. Accurate and efficient assessment of plant diversity is critical for guiding desert steppe restoration efforts. However, desert steppe vegetation has sparse leaves and sparse distribution. It is difficult to accurately distinguish micro-vegetation types based on a single spectrum, vegetation index or texture feature, and the resolution of satellite remote sensing cannot meet the needs of high-precision diversity assessment. To this end, this study proposed a novel method for assessing plant diversity index in degraded desert grassland based on multimodal UAV hyperspectral data and Encoder-CNN. Through experiments on different modal feature combinations, spatial spectra, vegetation indices and texture features were targeted and fused. Channel Attention Fusion (CAF) was introduced into Encoder to achieve cross-layer "soft" residual fusion, the Encoder and CNN models were fused to construct a global-local co-expression structure, and finally the quantitative calculation of the plant diversity index at the pixel level was realized. The results show that the vegetation types determined by the fusion of multimodal data and deep learning are consistent with the existing species, dominant species and sub-dominant species of the actual community, and the calculated diversity index results are also consistent with the actual situation. The use of multimodal data combining spatial spectral features with index features, combined with the Encode-CNN model, can provide the most accurate information on community composition. The overall accuracy of sparse vegetation classification can reach 90.01%, and the average accuracy can reach 85.23%, which is better than single mode or traditional 3DCNN, VIT models. This study demonstrates the application potential of UAV hyperspectral multimodal technology and deep learning in the assessment of desert steppe plant diversity, providing important technical support for ecological protection and conservation.