Abstract:Vegetation indices can reflect the spectral characteristics of different ecosystems and render remote sensing images easier to interpret, which are widely used to identify the ecosystem distribution patterns. A vegetation index is a specific expression for describing green vegetation, and its effects are inconsistent in different environments. The selection of a vegetation index needs to be combined with the characteristics of the application environment. Currently, selection of vegetation index is mainly based on the physical meaning of a vegetation index, which disregards its adaptability in the study area and leads to inconsistent research results. The correlation coefficients between vegetation indices were integrated into the vegetation indices selection algorithm based on Mahalanobis distance, and then a decision tree model was constructed based on the most suitable vegetation indices of the research area which were determined according to the selected samples. Taking Yongfeng County of Jiangxi Province as an example, it was attempted to identity the distribution pattern of ecosystems. Using this method, the ecosystems needed to be extracted were firstly determined and the relationship between ecosystems and decision tree nodes was established. Then, six different surface features, including wetlands, forests, grasslands, farmlands, urban and bare land, were classified. The overall accuracy of identification by the method was 89.11%, which was higher than that of the traditional methods. Taking wetlands as an example, the classification accuracy of vegetation indices determined by the vegetation selection algorithm was 91.62%, which was higher than the common vegetation indices that had an accuracy of 87.60%. The results indicated that the vegetation indices selection algorithm developed was applicable and effective. The method was a valuable and applicable tool for the extraction of regional ecosystem types and ecosystem management.