Method And Application Of Improving The Accuracy Of Land Cover Classification Based On Terrain Factors

Land cover is real reflection of the earth's surface,and land cover change is an important part of global environmental change.With development of ability of human to use and remould land,land cover is being constantly changed.Global environmental problems are also becoming increasingly prominent,such as global warming,forest reduction,biodiversity reduction,soil erosion and desertification.People gradually realize that land use/cover change,which is closely related to human activities,is an important cause of global environmental change.Accurate land cover data is the key information source for studying global environmental change.It provides important data support for studying global change,such for water circulation,energy flow and landscape structure on the earth's surface.It is also the basis of research of land cover change.In order to obtain more accurate land cover information,many researchers did extensive research on methods to improve classification accuracy of land cover.At present,the methods to improve classification accuracy of land cover focus on changing classification methods,selecting better combination of classification features and integrating multi-source data,and there have been many related studies and research cases.However,land cover products(even some influential global land cover products)have different classification accuracy in different regions.Classification accuracy of land cover products is relatively high in some areas but low in other areas,which limits application of these products.Therefore,it is one of effective methods to improve precision of land cover classification by enhancing precision of the areas with low precision and preserving the areas with high precision,but there few studies in this field.In this paper,the Beijing-Tianjin-Hebei region was taken as the study area,land cover data at 1:250,000 was used for reference,and MODIS land cover product(MCD12Q1)was taken as a research object.Multiple linear regression models between land cover distribution and topographic factors were established to improve classification precision of MCD12Q1 in the regions with low precision.The improved results were analyzed and compared wirh original MCD12Q1 and the reference data to confirm effect and degree of classification improvement.The main contents and conclusions are as follows:(1)The relationship between land cover distribution and terrain factors was analyzed macroscopically.Five terrain factors(altitude,slope,aspect,topographic relief and surface cutting depth)were selected and graded to analyze the relationship between terrain factors and land cover area.The results showed that altitude,slope,topographic relief and surface cutting depth had significant impacts on spatial distribution of three land cover types(forest land,grassland and cultivated land),but aspect insignificant.(2)The multiple linear regression models between land cover distribution and terrain factors were built.Average elevation,average slope,average aspect,terrain relief and surface cutting depth of each county were extracted with county-level administrative units in the study area as statistical analysis samples.Multivariate linear stepwise regression analyses were conducted with topographic factors as independent variables and proportion of each land cover type to total area as the dependent variable.The multivariate linear regression models between land cover distribution and terrain factors were built based on regression analysis results.Two groups of models were constructed according that whether constant term was set to zero or not.(3)The fitted results of two groups of models were analyzed by the four indexes with land cover dataset at 1:250000 as reference data,which included proportion of each land cover category to total area of the study area,correlation between calculated and actual area proportion of each county,spatial consistency and confusion matrix.The analyzed results showed that:(a)The calculated area ratio for woodland had better fit results on three indexes(proportion of woodland area composition,the correlation between calculated and actual area proportion of each county,overall accuracy)when constant term was not set to zero.(b)The calculated result for grassland was also better on two indexes(tcorrelation between calculatet and actual area proportion,overall accuracy)when constant term was not set to zero.(c)Spatial consistency and overall accuracy of cultivated land by regression calculation were better(72.64% and 62.48% respectively)when constant term was set to zero.(d)The calculated result for artificial land with constant term set to zero and that for other land cover types with constant term not set to zero had bad effect.(4)MODIS land cover product(MCD12Q1)was improved based on the relationship between land cover distribution and terrain factors.Two improvement schemes were designed:(a)on the basis of integrated and unified classification systems,only three(forest land,grassland,and arable land)of all five land cover types of MCD12Q1 were improved to generate a new land cover product MCD-NEW3.(b)All five land types of MCD12Q1 were improved to generate a new land cover product MCD-NEW5.Similarly,land cover data at 1:250000 was used for reference data to evaluate MCD-NEW3,MCD-NEW5 and MCD12Q1 by area composition ratio,spatial consistency and confusion matrix.The results showed that:(a)compared with MCD12Q1,area proportion of woodland and grassland had been better improved in MCDNEW3 and MCD-NEW5.(b)Overall accuracy and Kappa coefficient of MCD-NEW3 increased by 12.51% and 0.15,and spatial consistency of woodland,grassland and arable land increased by 8.55%,27.44% and 7.23%,respectively.(c)Compared with MCD12Q1,overall accuracy and Kappa coefficient of MCD-NEW5 increased by 17.14% and 0.25.Generally,spatial consistency,producer's accuracy and user's accuracy of all land types were improved in MCD-NEW5.(d)MCD-NEW3 and MCD-NEW5 had higher classification accuracy compared with MCD12Q1.(5)Comparison of the two improved products.It was found from comparison of MCD-NEW3 and MCD-NEW5 that overall accuracy increased from 73.34% to 77.97% from MCD-NEW3 to MCD-NEW5,Kappa coefficient increased from 0.54 to 0.64,Producer's accuracy of four land cover types(woodland,grassland,artificial land and other land types)increased by 16.99%,14.38%,2.2% and 5.62% respectively,and user's accuracy of cultivated land,artificial land and other land types increased by 7.39%,25.21%,23.47%,but that of grassland increased by a small margin.Compared with MCD-NEW5,producer's accuracy of cultivated land and user's accuracy of woodland of MCD-NEW3 were higher.Generally,precision and improvement effect of MCD-NEW5 was better that MCD-NEW3.In summary,in this paper correlation between land cover distribution and terrain factors was analyzed to build multivariate linear regression models.Two schemes were designed to improve classification accuracy of MODIS land cover product based on these models.The goal of improving classification accuracy of MODIS land cover product has been achieved.The research method used in this paper not only improved classification accuracy of MODIS land cover product but also has reference value for studies on land cover classification in the future.