首页 > , Vol. , Issue () : -
Objective：Plant community plays an important role in wetland elemets and is vulnerable to human activities and climate change. Wetland plant communities classification and mapping provide scientific important data support for wetland ecological monitoring and evaluation. This paper aims to develop a classification scheme suitable for the wetland plant communities in Poyang Lake wetland. Method：Taking Poyang Lake National Nature Reserve as the research area, based on the monthly Sentinel-1 and Sentinel-2 time series data in 2019, this study extracts five types of image feature parameters, including water and vegetation indexes group, red edge indexes group, texture features group, spectral features group and polarization radar backscatters group, with a total of 240 feature indexes, and uses random forest, support vector machine and deep neural networks (DNN) algorithms for classification. To explore a set of optimal features combination and a suitable classification scheme for wetland vegetation mapping in Poyang Lake. Result：The results show that: (1) Compared with radar data, the extraction accuracy of optical data is significantly better than that of radar data in wetland plant communities classification and mapping. . Radar data can be used as a supplement to optical data when optical data is insufficient. (2) Screening the importance of each image features of Sentinel-2 helps to improve the classification accuracy. The preferred time periods are mainly distributed in January, May, August, September, October and December; (3) Five groups of unitary image features are selected to classify separately, the classification accuracy is as follows: red edge indexes group > water and vegetation indexes group > spectral features group > radar polarization data group> texture features group; (4) Comparing the combined image features groups with the unitary image features groups, the combined image features group are not necessarily helpful to improve the classification accuracy. The classification accuracy is as follows: red edge indexes group > water vegetation indexes group > combined image features group. Among them, the overall accuracy of the classification scheme using red edge indexes group and random forest method is 0.81, Kappa coefficient is 0.76; (4) Comparing the three classification algorithms, the classification accuracy is ranked as follows: Deep Neural Networks (DNN) > Random Forest (RF) > support vector machine(SVM). The overall accuracy of deep learning method does not greatly improve, which is only 2% higher than the random forest algorithm. Thus, Both of the deep learning method (DNN) and machine learning method (RF) can be used as the optimization algorithms. Conclusion：In conclusion, A classification scheme for wetland plant communities in Poyang lake wetland was proposed in this study using multi time-series Sentinel-2 and Sentinel-1 data. The optimal acquisition time periods of satellite data are in January, April, August, September, October and December. The optimal image features group can be red edge indexes group or water and vegetation indexes group for feature selection. The classification algorithm can select deep learning or random forest algorithm to classify wetland plant communities according to the requirements. This classification scheme can effectively improve the accuracy of wetland vegetation mapping in Poyang Lake, and provide scientific and technical solutions for decision-making departments.