Optimal Features Selection for Wetlands Classification Using Landsat Time Series

Accurate wetlands distribution maps could provide important information for wetland management and protection. While, wetlands have significantly inner-annual variation which makes accurate wetland mapping challenging. Time series data with high temporal resolution, such as MODIS are wildly used in monitoring wetlands, but the coarse resolution cannot precisely identify detail of the wetlands and smaller wetlands. In addition, optimal features for wetlands classification is still unclear. In this study, Landsat monthly composited time series were used to identify wetlands. Firstly, we generated the Landsat 30m land surface reflectance, NDVI, NDWI and TC-Wetness time series by compositing the Landsat7 ETM+ and Landsat8 OLI data. Then the extension of the Jeffries-Matusita Distance (J Bh ) and the Random Forest (RF) algorithm were used to select the optimal features for accurate wetlands classification. Finally, the optimal features were used for wetlands classification based on the RF algorithm. The results showed that: (1) the optimal features included NDVI in May and August, TC-Wetness in May and NDWI in June and August which can effectively separate seasonal and permanent marshes, rice fields, grass and drylands; (2) the overall accuracy and Kappa coefficient calculated using the optimal features were 90.05% and 0.88 respectively which were similar to the wetlands classification result derived by the entire Landsat time series data. The optimal features could improve efficiency and decrease data redundancy.