Study On Improving The Accuracy Of Land Cover Products Based On Terrain Factors Modeling

Land use/land cover data is one of important data sources for study on global ecological environment change.The precision of land use/land cover data products has been explored by most researchers.However,existing land cover products have a shortage of low precision at temporal resolution or spatial resolution.Therefore,how to improve the accuracy of existing land cover data products has become a hot topic.At present,many researchers have published research results on methods to improve the accuracy of land cover products,such as optimizing classification characteristics,changing classification method,and multi-source data fusion.However,land cover is significantly affected by natural factors,especially topographic factors.Topographic factors lead to different precision of land cover products in different topographic regions.Therefore,it may be an effective method to improve the precision of land cover products by establishing the relationship models between terrain factors and land cover.This study was supported by the Natural Science Foundation of Hebei Province of China – Study on Methods of Improving the Precision of Land Cover Classification based on Terrain Factors and Time-series NDVI.In this paper,Beijing-Tianjin-Hebei region was taken as study area,MODIS land cover products(MCD12Q1)as research object and 1:250000 land cover data(LC-TYPE)asreference data.Firstly,classification systems of the two land cover data were merged into five categories(forest land,grassland,cultivated land,construction land and the other).Secondly,MCD12Q1 was overlaid with LC-TYPE,and the overlaid result was divided into two parts.One is those regions or cells where there are consistent land cover categories for MCD12Q1 and LC-TYPE(called Consistent Areas);the other is other regions or cells where there are inconsistent land cover categories(called Inconsistent Areas).Thirdly,the whole study area was classified into plain area and non-plain area based on DEM data.Fourthly,five terrain factors(elevation,slope,aspect,surface cutting depth and relief degree)were used as dependent variables and area ratio of land cover categories as the independent variable to construct multiple regression models.Finally,the constructed models were applied to improve the precision of the MODIS land cover products.The main research results and conclusions are as follows:(1)The study area was divided into plain areas and non-plain areas.The five terrain factors were used as independent variables and area percentage of each land cover category to total area of five types of land cover at county level as the dependent variable.SPSS software was used to build multiple linear regression models(two situations: equations with a constant or without a constant).According to fitting accuracy of the models,the models without constants for plain areas and non-plain areas were chosen to improve classification accuracy of the MODIS land cover product.Two detail methods were designed:(a)the model without a constant for plain areas was applied to improve the classification accuracy of forest land,grassland,arable land,construction land,other land types;(b)the model without a constant for non-plain areas was applied to improve the accuracy of all land cover.With 1: 250000 land cover data as the reference,the precision of the improved land cover product was evaluated by using proportion of each land area and ratio of spatial consistency as evaluation indexes.The results showed that the spatial composition ratio of cultivated land,woodland and grassland for the improved land cover product(NEW-1-2)was closer to the reference data LC-TYPE than before it was improved,while the improved effect for construction land and other land category was not obvious.According to the spatial consistency ratio,the cultivated land has the best improvement effect,followed by grassland and woodland.The spatial consistency ratio of cultivated land between NEW-1-2 and LC-TYPE is 85.19%,and the ratios ofgrassland and woodland are 78.81% and 60.60%,respectively.Therefore,compared with the reference data,the spatial consistency ratio of the improved land cover product(NEW-1-2)is higher than that of the original MCD12Q1.(2)The models were constructed only based on data from Consistent Areas.With the terrain factors in Consistent Area as independent variables,the area percentage of each type of land cover to total area of five types of land cover in Consistent Area as the dependent variable,SPSS software was used to construct multiple linear regression models with constants and without constants,respectively.Finally,higher precision models without constants for forest land,grassland,cultivated land were used to improved MODIS land cover products.With 1: 250000 land cover data(LC-TYPE)as the reference data,the precision of the improved land cover product was evaluated by using the area proportion and the ratio of spatial consistency as evaluation indexes.The results showed that the proportion of forestland,grassland and cultivated land for the improved land cover product is closer to 1: 250000 reference data than before improvement.Among various kinds of land cover,the improvement effect of grassland was the most obvious.The difference was reduced to 0.03 from 0.21 before the improvement.According to the ratio of spatial consistency,the improvement effect of cultivated land is the best,followed by grassland and woodland.The ratio of spatial consistency for cultivated land between the improved product(NEW-2-2)and LC-TYPE reaches 85.40%,and those for grassland and woodland are 80.68% and 66.63%,respectively.A conclusion can be drawn that the models constructed based on Consistent Areas can further improve classification precision of land cover products that the models based on the whole study area.(3)It is demonstrated from comparison between the two groups of data in the above that effect of models based on Consistent Areas is better than that based on the whole study area.The models between distributions of land cover categories and terrain factors generally express quantitative relationship between them.Accuracy and reliability of the models depend on accuracy and reliability of data.For Consistent Areas,there are the same land cover categories in the same spatial positions for the two sets of data,so probability of correct classification in Consistent Areas is larger than that in other areas.This is exactly an explanation why effect of models based on Consistent Areas is better than that based on the whole study area.