Incorporating ancillary data into Landsat 8 image classification process: a case study in Hoa Binh, Vietnam

Abstract

This was the first study to assess improvements in accuracy related to ancillary data integration in Landsat 8 image classification since its launch in February 2013. Hoa Binh (northern Vietnam) is a mountainous province with natural forests at high elevations and planted forests on lower slopes. This study integrated a normalized difference vegetation index (NDVI) and digital elevation model (DEM) with the spectral bands of a Landsat 8 image to minimize the influence of shadows on image classification, distinguish between natural and planted forests, and produce a land cover map of Hoa Binh Province for forest inventory support. The image was geo-referenced to the projection of Vietnam (VN-2000) and digital numbers of bands 4 and 5 were converted to reflectance for the NDVI calculation. A DEM was generated from 1:50,000 topographic maps with 40-m contour intervals. A classification of accuracy was performed on a multisource dataset (bands 1–7, and 9, NDVI, and DEM) in comparison with results from a spectral image. The results indicated that user and producer accuracies increased by 14.36 and 11.29 % (natural forest), 7.27 and 10.33 % (regenerated forest), and 8.43 and 11.28 % (planted forest), respectively. Accuracies of identification of barren and agricultural lands, settlements, water bodies, and other classes increased insignificantly. Generally, overall accuracy improved by 5.23 % (from 84.51 to 89.74 %), and the kappa coefficient of the spectral classification was 0.72 compared with 0.86 for the ancillary classification. This study concluded that integration of DEM and NDVI data improved the accuracy of Landsat 8 image classification.