Abstract:
Nine key phase GF-1 WFV images covering the growing period of winter wheat from October 2015 to June 2016 were selected as data sources.The multi-temporal spectral characteristics, vegetation index and texture features were integrated to set up four groups of feature combinations. Then，in order to improve the accuracy of classifying the land type and extracting winter wheat planting area in Xuchang City, Henan province, features were selected according to the feature importance scores and the optimal feature subsets were obtained to establish a random forest classification model. The results showed that in the four feature combinations without feature selection, the combination of all features has the highest classification accuracy. After the feature selection, the classification accuracy of each group was improved, indicating that comprehensive multiple types of feature variables and feature selection can effectively improve the classification accuracy. The contribution rates of different feature types and different time series feature variables are different, among which vegetation index>spectral feature>texture feature. Compared to other months during the winter wheat growth period, February, March, May and June has a higher contribution rate.The total area of winter wheat in Xuchang City, Henan province is 2 258.7 km2.The overall classification accuracy reached 95.18%, the Kappa coefficient was 0.925 5, and the mapping accuracy and user accuracy of winter wheat reached 98.67%.
