Response of Spatial Scale for Land Cover Classification of Remote Sensing
Kaijian XU,Qingjiu TIAN,Yanjun YANG,Nianxu XU
2018-01-01
Abstract:Classification based on remote sensing data has been widely applied in land cover mapping and the dynamic change monitoring research, of which the consequence is always strongly affected by spatial resolution of the used images. However, the response of multi-resolution images to remote sensing classification is still highly uncertain. Satellite observation could supply more and more multi-resolution images covering the same area at the same time and it would provide abundant data and technical support for study of remote sensing classification. In this study, the Hetian basin of Changting County in Fujian Province, was selected as a case to examine the performance of three typical classifiers (Maximum Likelihood Classification, MLC;Support Vector Machine, SVM;Artificial Neural Network, ANN). They were applied to satellite observations of temporal quasi-synchronous and multi-spatial resolution from medium to high spatial resolution (1～50 m) and we investigated the links between spatial resolution and remote sensing classification. Then, we also analyzed the spatial scale difference of spectrum reflectance, recognition accuracy and area extraction of five major land types (including arable land, forest land, water area, bare land and construction land) of the data with seven spatial resolution levels of 1, 2, 4, 8, 16, 30, and 50 m. They were supported with GF-1 PMS (pan and multi-spectra sensor), GF-2 PMS, GF-1 WFV (wide field view), Landsat-8 OLI (operational land imager) and GF-4 PMS data. 1845 recorded points observed in field survey were taken as training samples and validation samples. The results showed that along with the change of image spatial resolution from 1 to 50 m, (1) the mean spectra of bare land and construction land remained stable and no obvious changes occurred to water body, while the mean spectra of arable land and forest land decreased significantly when image resolution coarser than 4 m. The standard deviations of water body, bare land and construction land all increased constantly, while the standard deviations of arable land and forest land almost maintained stable. (2) The overall accuracy gradually decreased from 94.97± 2.5％ to 79.03 ± 2.25％ across the three classifiers, showing a gradually downward trend. Meanwhile, Kappa coefficient also gradually decreased from 0.93 ± 0.03 to 0.72 ± 0.03, which indicated that the accuracy of land cover classification was closely and sensitively related to the resolution of remote sensing images (P<0.05). (3) The calculation errors of the land types area would become larger as the image tend to be coarser, of which the area of arable land, bare land and construction land decreased significantly, the area of forest land increased, and the change of water body was not evident. The results above confirmed that when using multi-resolution images to generate land cover area or making area comparison refer to time serial data results, the errors from spatial database of various multi-scale could not be neglected, which would be more suitable to make the multi-scale transform for spatial effect correction. Our framework demonstrated the regular pattern of multiscale remote sensing classification and provided the prerequisite for scale conversion of classification products with different resolution in the future.