Extract of land use/cover information based on HJ satellites data and object-oriented classification

Abstract: Land use/cover information is the basis for the study of regional climate and environment, and the information security of land resources planning and management. However, the accuracy of urban areas land use/cover information extraction is significantly affected by the high heterogeneity of land surface. Remote sensing technology, providing large-scale, timely, continuous and comprehensive measurements, has become an important means of land use/cover information extraction. HJ satellites, with high temporal and spatial resolution and wide coverage, provide a new way to fast and accurately extract large-scale land use/cover information. They have been widely used in land use/cover classification, crop information extraction, wetlands information extraction, and so on. The object-oriented classification method, which makes full use of the spectral information of remote sensing images and takes into account the spatial distribution characteristics and correlations of geographical objects, can compensate for the deficiency of traditional pixel-based classification methods. This study developed a supervised classification method for regional land use based on the object-oriented random trees algorithm to quickly extract land surface information with low cost and high precision. We selected the Changsha-Zhuzhou-Xiangtan core area as the study area and used the multi-temporal and multi-spectral information of HJ satellite CCD (charge-coupled device) data. Firstly, high quality HJ-CCD data (10 phases in total) were selected, and preprocessed by radiometric calibration, atmospheric correction, accurate geometric correction and image registration. The time series of normalized difference vegetation index (NDVI) and of the first principal component (PC1) were calculated, and their results overlapped each other. The best time series of HJ classification data were determined by the J-M (Jeffries-Matusita) distance variable separability analysis, combined with land cover of study area, and phenotypic characteristics difference of different vegetation. HJ data of the February, May, July, September, October and December phases are the best data combinations for land use/over information extraction in this study. Then the e-Cognition's multi-scale segmentation algorithm was employed to segment the HJ-NDVI, HJ-PC1, HJ-PC2 (the second principal component) of the best time series combination. The urban land use/cover information was classified by the object-oriented random forest algorithm. Finally, the accuracy of the algorithm was evaluated, and compared with that of the time series pixel-based classification and single-phase object-oriented classification. The results indicate that the land use/cover information extracted by the object-oriented classification method using time series HJ data is consistent with the real situation on range and distribution of each land type, and with less speckle noise. The overall accuracy and Kappa coefficient of this method are 91.55% and 0.90 respectively. Specifically, the accuracy is higher than 90% for the paddy field, irrigated land, dry land and forest, and is close to 90% for building land. Compared with the time-series pixel-based classification and single-phase object-oriented classification methods, the overall classification accuracy and Kappa coefficient of the proposed method are increased by 2.26%, 0.02 and 6.82%, 0.08 respectively. This means the best time series HJ combination data can fully utilize the seasonal spectral differences of different vegetation types, which avoid the spectral similarity among different vegetations in single-phase image. So the proposed method can effectively improve the accuracy of land use/cover information extraction in urban areas.