Land use/cover classification based on combining spectral mixture analysis model and object-oriented method

Complex land use/cover and fragmented land objects have posed a great restriction on the efficiency and accuracy of classification. In traditional classification, a single pixel was often taken as the basic unit, inevitably leading to the low accuracy of the mixed pixels. Thus, the low classification accuracy of land use/cover can be attributed that the mixed image pixels blur the spectral information of land objects. Meanwhile, it is necessary to efficiently utilize the spectral, shape and texture characteristics of land objects during extraction. In an object-oriented model, the adjacent pixels are taken as the objects considering various attributes, such as spectrum, shape and texture, in order to weaken the interference of mix pixels to land use information extraction. However, a large number of feature parameters in the object information extraction can reduce the computational efficiency and classification accuracy. As a result, it is highly demanding for the combined technology to realize the automatic and high-precision land use/cover classification using remote sensing images. In this study, a land use/cover extraction was carried out to integrate the spectral mixture analysis and object-oriented model using the Sentinel-2A images, in order to improve the accuracy of land use/cover classification. Firstly, the rules for land object extraction were constructed by 8 characteristic parameters, such as Normalized Difference Water Index (NDWI), Normalized Difference Vegetation Index (NDVI), and Soil Background Level (SBL) at three optimal segmentations, according to the spectral, shape and texture of land objects. Secondly, the spectral mixing was utilized to extract the three generic endmembers in the study area, including substrate (SU, rock and soil), green vegetation (GV, photosynthetic leaves), and dark material (DA, shadow and water). Finally, an illustration was presented for the effects of spectral features of three endmembers on the optimization of extraction. The results showed that: 1) The overall accuracy of land use/cover classification was 80.83% for five land objects using the fuzzy function and threshold in different hierarchical levels, where the Kappa coefficient was 0.76. 2) The spectral extraction significantly improved the overall accuracy of land use/cover classification up to 90.00% using the fusion of three endmembers derived from spectral mixture, where the Kappa coefficient was up to 0.88. 3) The integration of three endmembers with clear physical meaning enhanced the difference of each component in the pixel, especially in cultivated land and construction land. Correspondingly, the deficiency was reduced for traditional spectral indexes in the resolution between vegetation and soil brightness, due mainly to the explicit physical meaning of three endmembers. Besides, this model was conducted from easy decreasing, thereby to decrease uncertain factors layer by layer. Thus, it is also expected to make full use of spectral features, suitable for the medium and high resolution of remote sensing images with multiple spectral bands. The finding can provide great potential to the fine extraction for land use information.