| Land-cover information is indispensable for ecological environment assessment,national-condition monitoring and macro-control analysis,and also plays important role in biodiversity protection,urban-rural pattern planning and natural resource management.Recently,with the developments of remote sensing and compute techniques,it was of great significance to monitor the status and trends of global land-cover for scientifically understanding the characteristics and distributions of different land-cover types over the Earth.Over past years,several global land-cover products,covering different spatial resolution,temporal updating frequency and classification system,have been released,however,these is still a lack of the global land-cover product simultaneously containing high spatial resolution and fine classification system.In addition,as it usually takes a lot of manpower and computer resources to collect global coverage training samples and process the mass remote sensing imagery,which makes the global land-cover mapping still suffering great difficulties and challenges.Aiming at mapping the global fine land-cover at 30-m using quantitative remote sensing techniques,several specific studies,including quantitative processing of time-series Landsat imagery,Data Cube tiling re-management,development of the Global Spatial-Temporal Spectral Library(GSPECLib),global impervious mapping at 30-m and fine land-cover mapping and validation over China or Globe using the SPECLib,were carried out one after another.The main results are listed as following:(1)The quantitative processing of Landsat data is an important basis for automatic classification of land-cover.In this paper,the terrain radiation C correction model and atmospheric correction algorithm were developed to eliminate the effects of terrain and atmosphere radiation transfer.Then,a novel multi-temporal cloud and shadow detection algorithm(MTICZ)is proposed,which significantly improves thedetection accuracy of clouds and shadows,which has been validated to achieve an improvement of about 6% compared to the Fmask algorithm.Lastly,a new Landsat Data Cube management mode is designed,which significantly improves the data usage efficiency of the track overlap areas.(2)The combination of multi-source remote sensing datasets could significantly improve the mapping accuracy of global impervious surfaces products.Specifically,due to the extremely complicated spectral and spatial heterogeneity of impervious surfaces,it is difficult to guarantee the mapping accuracy of the impervious surface only using multi-temporal spectral information.In this study,the combination of multi-source and multi-temporal remote sensing datasets(Landsat SR,Sentinel-1 SAR,VIIRS NTL,SRTM topographical variables and Globe Land30land-cover product)is proposed to automatically produce the global 30 m impervious surface products in 2015,which is validated to achieve an overall accuracy of 96.7% and a kappa coefficient of 0.903 and outperformed other products especially for broken impervious surfaces.(3)Using the prior spectra in the GSPECLib,it is possible to automatically realize large-area land-cover classification.In this study,the time-series reflectance spectra in the GSPECLib perform great ability to replace the training samples in the supervised classification.Using the reflectance spectra in the GSPECLib,time-series Landsat Data Cube SR and decision-level fusion classification model,the annual 30 m land-cover product with fine classification system in China is successfully produced,which has been validated to achieve an overall accuracy of 71.3% and80.7% and the kappa coefficient of 0.664 and 0.757 for the level-2 validation system(19 land-cover types)and the level-1 validation system(nine land-cover types),respectively.(4)Based on the GSPECLib and Google Earth Engine cloud-computation platform,it is possible to accurately and efficiently produce global 30 m land-cover products with fine classification system.Specifically,although the reflectance spectra in the GSPECLib have ability to produce regional land-cover classification,it usually suffers low mapping efficiency because of excessive needs of computing and storage resources.In order to significantly improve the mapping efficiency especially for global land-cover mapping,the training samples in the GSPECLib,the time-series Landsat SR imagery and locally adaptive random forest classification model are combined to produce the global 30-m land-cover products with fine classification system in 2015(GLC_FCS30-2015),which has been validated to achieve an overall accuracy of 81.4% and a kappa coefficient of 0.772 for the level-0validation system(ten land-cover types),an overall accuracy of 70.8% and kappa coefficient of 0.678 for the UN-LCCS(United Nations Land Cover Classification System)level-1 system,and an overall accuracy of 68.1% and kappa coefficient of0.656 for the UN-LCCS level-2 system.The main innovative contributions of this study contain:(1)An automatic processing framework,developed for quantitative processing of time-series Landsat imagery and global Landsat Data Cube,performs great ability to simultaneously eliminate the effects of terrain and atmosphere radiation transfer,improves the detection accuracy of identifying cloud and shadow pixels,and increases the data usage efficiency of the track overlap areas using the Data Cube mode.(2)A novel and automatic mapping method for impervious surface is proposed by combining multi-source and multi-temporal remote sensing datasets,and a global30 m impervious surface product in 2015 has been produced.The validation results indicated that it improves the mapping accuracy especially for complicated impervious surfaces(such as: rural buildings and road networks).(3)A novel quantitative remote sensing classification strategy using the prior GSPECLib has been proposed to solve the problem of huge human intervention for large-scale land cover mapping.For the first time,the global 30 m land-cover products with fine classification system(GLC_FCS30)has been accurately generated using the GSPECLib-based method. |
PDF Full Text Request