Objectives The monitoring of salt marsh vegetation is the basis for protection and restoration of ecological functions in the Yellow River Delta (YRD) wetland. Due to the complexity of vegetation growth conditions and the spread of the invasive species Spartina alterniflora, vegetation detection and classification are particularly important.

Methods This paper takes part of the YRD wetland as the study area, and chooses high-resolution aerial images as data source. Four feature variables are generated, including spectral features, color features, index features and texture features. Six different classification schemes are constructed. Scheme 1 with only spectral features is regarded as the control group. Index features, texture features and color features are integrated into scheme 2, scheme 3 and scheme 4, respectively. Scheme 5 contains all features, and scheme 6 constructs a multi-feature optimization feature set. Random forest method is used to classify vegetation for each extraction scheme and the corresponding accuracies are verified, aiming to explore the influences and reasons of different feature variables on the classification. The best preferred features are selected to improve the effect of vegetation classification.

Results Based on the visible and near-infrared spectra, just adding different features to the experiment has different effects on the accuracies of vegetation classification. Scheme 1, scheme 2 and scheme 3 have unsatisfactory extraction effect on Phragmites australis and Suaeda salsa. Scheme 4, scheme 5 and scheme 6 with the addition of color features can better distinguish between the two, probably because Suaeda salsa shows dark red on the image, which is quite different from other vegetation. Scheme 4 divides non-vegetated tidal flat areas and water body edges into Phragmites australis, which are similar in color, and adds color features results in some misclassifications. Scheme 5 and scheme 6 are well classified, but due to the mixed nature of vegetation, all schemes have varying degrees of misclassification of Tamarix chinensis, Phragmites australis and Spartina alterniflora. Index features have positive effects on Suaeda salsa extraction, texture features reduce the accuracy of vegetation classification, and the integration of color features is the key to improve overall accuracy of classification. Based on multi-feature optimization of random forest, the extraction effect is the best, with overall accuracy of 88% and Kappa coefficient of 0.85.

Conclusions The main advantages of this study are the acquisition of new data sources, the introduction of multiple feature variables, and the experimental evaluation and classification accuracy analysis of different feature variables. The importance of color features is verified and multi-feature optimization of random forest by integrating color features is a feasible method to classify vegetation information in the YRD. The proposed method can effectively distinguish vegetation from non-vegetation and extract each vegetation type at the same time. This study provides an effective technical route in feature selection and methodology for vegetation information extraction in the YRD.