A comparison of pixel-based and object-based land cover classification methods in an arid/semi-arid environment of Northwestern China

Land cover classification provides useful information of vegetation for land surface models especially in arid areas. Due to the complex landscapes in the downstream area of the Heihe River basin in northwest China, accurate land cover classification plays a key role in better understanding of the eco-hydrological processes, identification of water consumption by different land cover types and effective use of scarce water resources in this region. The aim of this study is to map land cover classification in the arid environment in the downstream area of the Heihe River basin by employing both pixel-based and object-based image analysis methods using high spatial resolution data. The data used in this study were 2.5 m high spatial resolution imagery (SPOT-5) and field survey. Vegetation indices and land surface texture characteristics were calculated from SPOT-5 imagery and used as input variables for classifications. In this study, overall classification accuracies between pixel-based and object-based classification methods were statistically significant when the same Support Vector Machine algorithm was applied. Object-based classification methods could avoid the salt-and-pepper noise existed in the pixel-oriented results and achieved more accurate depictions of land cover types in this region than the results using pixel-based algorithms. The accuracy assessment on the classification of single tree crown of Populus euphratica revealed more promising results by using object-based image analysis than using pixel-based classification methods. The low accuracy in identifying grassland class was due to the inadequate selection of samples for classifications in the study area. Inappropriate scale value of segmentation resulted in the low producer´s accuracy of residential areas when using object-based algorithms, whereas the misclassification led to the low user´s accuracy of residential areas when utilizing per-pixel methods.