Finer Resolution Land-Cover Mapping Using Multiple Classifiers and Multisource Remotely Sensed Data in the Heihe River Basin

Land-cover datasets are crucial for research on eco-hydrological processes and earth system modeling. Many land-cover datasets have been derived from remote-sensing data. However, their spatial resolutions are usually low and their classification accuracy is not high enough, which are not well suited to the needs of land surface modeling. Consequently, a comprehensive method for monthly land-cover classification in the Heihe river basin (HRB) with high spatial resolution is developed. Moreover, the major crops in the HRB are also distinguished. The proposed method integrates multiple classifiers and multisource data. Three types of data including MODIS, HJ-1/CCD, and Landsat/TM and Google Earth images are used. Compared to single classifier, multiple classifiers including thresholding, support vector machine (SVM), object-based method, and time-series analysis are integrated to improve the accuracy of classification. All the data and classifiers are organized using a decision tree. Monthly land-cover maps of the HRB in 2013 with 30-m spatial resolution are made. A comprehensive validation shows great improvement in the accuracy. First, a visual comparison of the land-cover maps using the proposed method and standard SVM method shows the classification differences and the advantages of the proposed method. The confusion matrix is used to evaluate the classification accuracy, showing an overall classification accuracy of over 90% in the HRB, which is quite higher than previous approaches. Furthermore, a ground campaign was performed to evaluate the accuracy of crop classification and an overall accuracy of 84.09% for the crop classification was achieved.