Abstract: Wetland is an important transition zone between terrestrial and aquatic ecosystems. Accurate and efficient acquisition of wetland plant community distribution information is of profound significance for wetland conservation. We use the UAV multispectral images as the data source, construct a multidimensional feature dataset containing spectral features, vegetation indices, and texture features, and determine the optimal feature dataset by the Relief F algorithm. Then, we construct a convolutional neural networks (CNN) classification model based on the feature selection and classify the optimal feature dataset with the CNN based on the original multispectral images. The results show that:① The classification accuracy increases and then decreases as the number of features increases, and the classification accuracy is highest when the number of features is 32. ② Texture features such as information entropy, homogeneity, and multispectral vegetation indices such as GNDVI, MSAVI2, and RVI extracted by GLCM with a window of 13×13 have higher importance. ③ The CNN classification model based on the optimal feature dataset can effectively extract the spatial-spectral information and suppress the “salt-and-pepper noise”, with the best classification effect and the overall accuracy of 93.40%, which is 9.80% and 7.40% higher than the RF and CNN classification models without feature optimization, respectively.

Key words: UAV multispectral, convolutional neural network, feature preference, wetland fine classification