A land cover classification system for use in global change modeling and based on BISE algorithm

A Land Cover Classification System for Use in Global ChangeModeling and based on BISE AlgorithmQuanfang WANG∗a, Haiwen ZHANG a, Hangzhou SUN a, Jiayong LI ba Faculty of Resources and Environmental Science, Hubei University, Wuhan 430062,China;b Institute of Geographical Sciences & Natural Resources Research, Chinese Academy ofSciences, Beijing 100101, ChinaABSTRACTIn this paper, a new logic for land cover classification at regional scale has been introduced. The critical features of this classification are that: 1) indeed distinguished from land use to avoid the confusion between land use types and land cover types; 2) based on remote sensing so that repeatable and efficient re-classifications of existing land cover will be possible; 3) based on spectrum and primary attributes of plant-canopy structure, that are important to globe change modeling and can be measured in the field for validation or/and by remote sensing; 4) based on the phonological difference among broadly defined vegetation because some typical land cover is easily distinguished by using the characteristics of seasonal dynamic; 5) based on component and function properties (e.g., influence on land surface processes) of covers to differentiate mixed land cover. Following the above ideas and using time-series MODIS 250 m data (i.e. NDVI and reflectance) which were reprocessed by a BISE algorithm to identify contaminated pixels with residual cloud, a two-level land cover classification scheme was produced for the southeast Hubei Province and middle Qinling Mountains in Shanxi Province, China. Results show there were seven primary classes and fifteen sub-classes identified and mapped.Keywords: Land cover classification system, Time-series MODIS 250 m data, BISE Algorithm1.INTRODUCTIONAccurate information on land cover at regional to global scale is required by global modeling (e.g., climate models, carbon cycle models). However, it is very difficult to meet the demand in spite of abundant existing data on land cover and its changes. As for reasons, it mainly results from the absence of a standard land cover classification system with quantificational classification criteria so as to compare the data collected from many different sources for a specific purpose. Moreover, the absence of an unambiguous, repeatable definition of land cover [1] has also caused the confusion between land cover and land use. For example, land use and land cover were always regarded as a whole using “land use and land cover” or “land use/land cover (LULC) ” and land use types were employed for the replacements of land cover types [1-5]. As a result, it’s also nearly impossible to aggregate the present data on land cover.In this study, the objective is to introduce a new logic for land cover classification, which could help to solve the problems stated, and apply it to build a Land Cover Classification System (LCCS) at regional scale.2.STUDY AREA AND DATA PREPROCESSING2.1Surveys of study areaTwo areas between Yangtze River Basin and Weihe River Basin have been studied in this paper. The first one is the hilly southeastern Hubei Province (see Fig.1) in transitional zone from Mid-Subtropics to Northern Subtropics, where contradiction of more population and less cultivated land is comparatively outstanding and Yangtze River forms the division between the northern part with lower altitude and the southern part with higher altitude of the region. As for vegetation, there are mainly Evergreen Broadleaved Forest, Deciduous Broadleaved Forest, subtropical Coniferous forest, Mixed evergreen-deciduous Broadleaved Forests, Paddy, Wheat and Ramie[6].∗wangqf@; phone 86-27-8866-1099PIAGENG 2009: Intelligent Information, Control, and Communication Technologyfor Agricultural Engineering, edited by Honghua Tan, Qi Luo, Proc. of SPIE Vol. 7490, 749005© 2009 SPIE · CCC code: 0277-786X/09/$18 · doi: 10.1117/12.836419Another Study area is the middle Qinling Mountains in Shanxi Province in China (see Fig.2), which has been recognized as the geo-ecological boundary between Subtropical and Warm-temperate zones in eastern China [7, 8]. As a transitional ecotone, it has abundant plant species and a variety of vegetation types, such as evergreen coniferous forest, deciduous coniferous forest, deciduous broadleaf forest, shrub, meadow, mixed forests, and crops, etc. moreover, human activities (e.g. forest clear-cutting, selective logging, agricultural reclaiming, urbanization encroachment) and natural disturbance (e.g. fire, insects, climate warmer and drier) have resulted in substantial losses of old-growth forests and fragmentation in vegetation landscapes in the region [9-11]. For example, Qinling fir (Abies chensiensis), endemic to China and an endangered plant, which was listed in the China Plant Red Data Book as one of the third categories protected plants. Now it is only found scattered in small forest fragments in Mts. Qinling, Bashan and Shennongjia at the altitude of 1300 m to 2300 m and still tending to degenerate recently [11]. Therefore, the status of land cover in Qinling Mountains has received much attention. Yet, few earlier studies have used remote sensing data to document the areas and the types, which mainly results from high spatial complexity of land cover there, and a feasible method of regional-scale land cover mapping based on remotely sensed data is challenging [5] because it requires those data to have large geographic coverage, high temporal resolution, adequate spatial resolution relative to the typical field size, and minimal cost.(a) Hilly southeastern Hubei Province in China (b) Middle Qinling Mountains in Shanxi Province, ChinaFig. 1. Elevation maps of two study areas, which were derived from the SRTM 90m Elevation Imagery.2.2Remote sensing data preprocessingA real, feasible method of land cover classification must be derived from remote sensing so that repeatable and efficient global re-classifications of existing land cover will be possible [1]. However, global-scale and Regional-scale land cover classification all requires remotely sensed data that have large geographic coverage, high temporal resolution, adequate spatial resolution relative to the typical field size, and minimal cost. Remotely sensed data from traditional sources such as the Landsat Thematic Mapper (TM and ETM+) and the Advanced Very High Resolution Radiometer [AVHRR) have proved useful for LULC characterization, but are limited for such an approach because of resolution limitations, data availability/quality issues, and/or data costs[5, 12,13].The MODIS instrument, a new generation of land surface monitoring, offers new possibilities for large-area land-cover mapping by providing a near-daily global coverage of science-quality, intermediate resolution (250 m) image, including vegetation index (VI) and reflectance data, since February 2000 at no cost to the end user[14]. In particular, MODIS 250 m NDVI data, are well suited for land-cover classification and monitoring because these data have sufficient spatial, spectral, temporal, and radiometric resolutions to detect unique multi-temporal VI signatures for some typical land cover such as forest, crop, grassland, etc [2, 15, 16]. According to the location of study area, 250m 16-day composite time-series NDVI and some reflectance imageries of MODIS (MOD13Q1) over the period of December 19, 2005 to January 3, 2007 were acquired. Then the tiled NDVI data were mosaicked, re-projected from the Sinusoidal to Albers Conical Equal Area (ACEA) projection and subset over the study area for each composite period and then sequentially stacked to produce the time-series data set.2.3Contaminated pixels with residual cloud in the time-series MODIS 250 m NDVI data identifiedBecause cloud contamination always reduces the accuracy of classification on remote sensing images, MODIS 250 m 16-day composite NDVI data provided by NASA are generated by selecting the pixels that have the maximum NDVI values within a 16-day period to minimize the effect of cloud cover and variability in atmospheric optical depth. But in the region with long-term cloudy or rainy weather, the maximum NDVI value composite (MVC) approach only eliminates most cloudy pixels [12], and some images perhaps contain residual cloud contamination. White et al. (1997), Reed et al. (1994) and Viovy et al. (1992) had developed the Best Index Slope Extraction (BISE) technique for identifying contaminated pixels with residual cloud and reducing the effects of cloud contamination and atmospheric interference in seasonal NDVI time series from daily AVHRR data [17-19]. The BISE algorithm is:1) to calculate the difference in NDVI values for pixel i between time t - 1 and time t, and the difference in NDVI between time t + 1 and t, respectively:dNDVI t-1,t = (NDVI t-1 – NDVI t) (1)dNDVI t,t+1 = (NDVI t+1–NDVI t) (2) 2) to adopt a threshold of 20% decline in the NDVI value at time t for identifying cloudy pixels. If dNVIt-1,t and dNDVIt,t+1 were at least a 20% decline from both NDVIt-1 and NDVIt+1, respectively, it is assumed that the pixel was affected by clouds, and NDVIt values were replaced by the averaged NDVI values at time t - 1 and time t + 1. We applied the algorithm to smooth the NDVI time-series data over the period of January 1, 2006 to December 18, 2006 (a total of 22 time periods). The NDVI data for the periods of December 19, 2005 to January 3, 2006 and December 19, 2006 to January 3, 2007 are the starting and ending points in the time series and therefore were not corrected and not used for image classification.3.DESIGNING A LCCS FOR USE IN GLOBAL CHANGE MODELING3.1Basic principleFive basic principles has been adopted in this paper, which is: ⅰ) to be distinguished from land use indeed and adhere to clear and strict class boundary definitions so as to avoid disagreement among authors of the existing areal extent of different land cover classes; ⅱ) to be based on simple, unambiguous, quantificational characteristics of land cover which are important to globe change modeling; ⅲ) to be based on remote sensing so that repeatable and efficient global re-classifications of existing land cover will be possible[1]; ⅳ) independent of the scale and adopted data sources [4]; ⅴ) open-end, i.e. the LCCS is a prior as well as flexible classification system, in which only the major classes are distinguished in advance and the sub-classes (attached to the major land cover class) will be defined by users adopting the combination of a set of independent diagnostic criteria so as to the prior classification system extended [4].3.2Classification criteriaSo far, multitudinous land cover classification systems have been already built [3,4,20-26] , but none of the current classifications has been internationally accepted, which is mainly caused by the land cover classes inappropriate for particular purposes (e.g., statistical or rural development needs), the scale related to a specific purpose or the confusion between land cover and land use, and some factors often used in the classification system that result in a undesirable mixture of potential and actual land cover (e.g., including climate as a parameter) [4].Because of the above, we suggest that a land cover classification system adopt the following criteria:1) spectrum (color), which is the first criterion of the classification and the most important basis for land cover classification. According to the difference in characterization of spectral response, it’s easy to distinguish green land cover (dominated by Vegetation), blue land cover (Inland water or Ocean), grey land cover (Semi-desert, Urban or Built-up land), white land cover (Snow and Ice), Non-Vegetated land cover (Desert), and so on;2) primary attributes of plant-canopy structure. Running et al [1] have shown that a complete globe vegetation classification should be derived from combinations of three primary attributes of plant-canopy structure, i.e. permanence of aboveground live biomass which defines whether the vegetation retains perennial or annual aboveground live biomass, leaf longevity (often termed evergreen versus deciduous canopy), and leaf type (needleleaved, broadleaved orgrasses). In fact, these primary attributes of plant- canopy structure can also be adopted as the classification criterion of land cover with its embedded information on vegetation [13], because vegetation is the primary component of land cover and covers most of the land;3) phenological difference among broadly defined vegetation. Some typical land cover is easily distinguished by using the characteristics of land cover with the change of seasons that can be derived from multi-temporal remote sensor data. For example, the Normalized Difference Vegetation Index (NDVI) of evergreen land cover is high and no remarkable change in one year, while deciduous forest has one peak of NDVI sliding curve in one year, annually double crops two peaks, and annually triple crops three peaks. Therefore, the classification criteria of land cover can be quantified with the eigenvalue showed by NDVI, such as greenness value, crest value, trough value, green-up period, green-end period, integral of NDVI (NDVI-I), difference between the maximum and minimum NDVI (NDVI-MM), and waveform of NDVI sliding curve in one year, etc;4) component and function properties of covers. Gopal et al have shown that land cover, perhaps the most fundamental characteristic of a place, is an integrated expression of the physical, climatic, and biotic environment as well as the history of land use by humans [13]. At the same time, its functions, particularly reflectivity and net discharge of CO2, vary with different component and structure of covers. So some mixed land cover is better differentiated by its constituent characteristics and influence on land surface processes.3.3Building the classification systemTo guide the interpretation of remote sensing data, a two-level, hierarchical classification scheme (see Table 1) was built by adopting a combination of the above diagnostic criteria and with the help of open factual data of land cover (such as books, papers/reports, and maps, etc.) on the hilly southeastern Hubei Province and the middle Qinling Mountains in China. At the initial stage (Level Ⅰ), the entire study area was mapped into seven classes, i.e. evergreen cover (woody), seasonal green cover (woody), seasonal green cover (herbaceous), seasonal green cover (crops), seasonal green cover (mixed), grey cover, and blue cover. The evergreen cover was then further classified into four sub-classes (Level Ⅱ).Table1. Hierarchical Land cover classification scheme for the hilly southeastern Hubei Province and the middle Qinling Mountains, ChinaCode and Name of basic classes Code and Name of sub-classes1 Evergreen cover (Woody) 11 Evergreen broadleaf forest12 Evergreen coniferous forest13 Evergreen shrub14 Evergreen mixed forest2 Seasonal green cover (Woody) 21 Deciduous broadleaf forest22 Deciduous coniferous forest23 Deciduous shrub24 Deciduous mixed forest3 Seasonal green cover (Herbaceous) 31 Grasses32 Wetland4 Seasonal green cover (Crops) 41 Single cropping in one year42 Continuous double cropping in one year43 Continuous triple cropping in two years5 Seasonal green cover (Mixed) 51 Mixed cover of crop and tree52 Mixed cover of crop, grass and shrub53 Mixed cover of crop and water6 Grey cover 61 Urban or Built-up land62 Barren or Sparsely land7 Blue Cover 71 River72 Lake73 Reservoir74 Other waterbodiesMoreover, each of those classes on level 1 is identified by simple, observable, unambiguous characteristics that are important to ecosystem biogeochemistry and could be measured in the field for validation. For instance, evergreen cover (Woody) primarily consists of woody plant (maybe natural forest, half natural forest and/or planted forest), keeps green all year, and has high greenness-degree index without obvious change in the whole year. Seasonal green cover (Woody) mainly constitutes with deciduous woody plant and its curve of greenness-degree index often has one obvious peak in a year, but the average greenness-degree index is lower than evergreen cover. Seasonal green cover (Herbaceous) is usually classified into two types including annual grasses and wetland, and also has one obvious peak in the curve of greenness-degree index in a year. In particular, wetland perhaps has extremely low greenness-degree index similar to water bodies when submerged. Grey cover usually has low greenness-degree index without obvious change in the whole year and may be classified into two types: 1) Barren or Sparsely land, such as the sand and the bare land, etc., which coverage degree of vegetation is throughout lower than 10% in a year; 2) Urban or Built-up land, which mainly consists of artificial buildings, may be mapped into cities, countryside inhabitant, the constructions uses, etc., and extremely similar to sand and other grey cover on spectrum characteristics. Blue cover, with extremely low greenness-degree index, is usually classified into two types, i.e. ocean and territorial water body (river, lake, and reservoir, etc.), but ocean is not applicable to the study area.4.MODIS-BASED IMPLEMENTATION AT REGIONAL-SCALE4.1Results of classificationIn this study, the supervised classification technique (maximum likelihood) was used to MODIS 250 m time-series NDVI imageries spanning one growing season, on which training areas were selected based on some MODIS 250 m 16-day composite reflectance imageries (MOD13Q1) in winter and the data of forest resource inventory provided by Forestry Department of Hubei Province and Shanxi Province in China. Finally, seven primary classes and fifteen sub-classes were identified and mapped for the two study area (see Fig. 2) by synthesizing the laws of distribution and succession processes of vegetation, identifying power on surface cover of MODIS 250 m NDVI data, the characteristics of forest resource inventory data, structure and functional properties of covers, etc.(a) Hilly southeastern Hubei Province in China (b) Middle Qinling Mountains in Shanxi Province, ChinaFig. 2. Land cover classification maps of two study areas derived from 250m 16-day composite time-series imageries of MODIS (MOD13Q1) over the period of December 19, 2005 to January 3, 20074.2Accuracy assessmentAccuracy assessment adopted Error Matrix, which is often used to measure the thematic classification accuracy and consists of classified data and reference data [27-30]. From the Error Matrix, many statistics value about accuracy can be produced, such as Overall Classification Accuracy (OCA) and Overall Kappa Statistics (Kappa).The result of accuracy assessment on the sub-classes showed that OCA of land cover classification in the study region equals to 75.32% and Kappa 67.01%. The interpretation accuracy of three types of seasonal green cover (i.e. deciduous broadleaf forest, evergreen broadleaf forest and crops) was over 80%, evergreen coniferous forest and wetland over 70%, while others had low accuracy (<60%). In particular, error classification often occurs among barren land, urban or built-up land and water. The reason is mainly from the phenomenon of different objects with same spectrum and/or same object with different spectrum, which often occurs when only spectral feature used to identify land cover.5.CONCLUSION AND DISCUSSIONLand cover mainly represents morphological property and dynamic characteristics of the earth's surface. Its accurate representation is a continuing challenge even though based on remote sensing. For global modeling requirements, a standard land cover classification system should adhere to quantificational classification criteria, an unambiguous, repeatable definition of land cover and clear boundary definitions between two classes. In this study, a simple classification logic for land cover was proposed and implemented in the hilly southeastern Hubei Province and the middle Qinling Mountains in Shanxi Province, China based on time-series MODIS 250 m NDVI data and some reflectance imageries in winter. The results shows that the new logic of land cover classification is feasible at the regional scale, but the accuracy needs to be improved by combine spectral characteristics with ancillary data including elevation data, shape feature of land cover (very useful to linear river identified), climate and eco-region variables as to eliminate confusion [1].ACKNOWLEDGEMENTST his study benefited from the NASA's Earth Observing System (EOS) Data Gateway for the MODIS NDVI products. And it was supported by Program of National Science Foundation of China (NO.40601003) and Science &Technology Research Program of Education Office of Hubei Province, China (NO.Q200610002).REFERENCES[1]S.W. Running, T.R. Loveland, and L.L. Pierce, “A vegetation classification logic based on remote sensing for use inglobal biogeochemical models”, AMBIO, 23(1), 77-81 (1994).[2] B.D. Wardlow, S.L. Egbert, and J.H. Kastens,“Analysis of time-series MODIS 250 m vegetation index data forcrop classification in the U.S. Central Great Plains, Remote Sensing of Environment”, 108, 290−310 (2007).[3]J.R. Anderson, E.E. Hardy, J.T. Roach, and R.E. Witmer, “A land use and land cover classification System for usewith remote sensor data”, U.S. Geological Survey Professional Paper 964. USGPO, Washington, D.C. (1976).[4]Antonio Di Gregorio, and Louisa J.M. Jansen, “Land cover classification system (LCCS): classification conceptsand user manual for software version 1.0”, Food and Agriculture Organization of the United Nations (FAO), Rome (2000).[5] B.D. Wardlow, and S.L. Egbert, “Large-area crop mapping using time-series MODIS 250 m NDVI data: Anassessment for the U.S. Central Great Plains”, Remote Sensing of Environment, doi:10.1016/j.rse.2007.07.019 (2007)[6]Yingming WANG, “On the vegetation regionalization of Hubei province,” Journal of Wuhan Botanical Research.3(2), 165-174 (1985). [in Chinese][7]S.E. Liu, “Botanical geography in North and West of China”, Contributions from the Institute of Botany NationalAcademy of Peiping, 2(9), 423-464 (1934). [in Chinese][8]Editorial Committee of Vegetation of China. Vegetation of China, Beijing: Science Press (1980). [in Chinese][9]M.D. Lei (Ed.). Vegetation of Shanxi Province. Beijing: Science Press (1999). [in Chinese][10]X.G. Zhao, and L. Xiao, “The vegetational and environmental history in Qinling Mountain since 15 ka BP”, ActaBotanica Boreali-Occidentalia Sinica, 23(4), 523-530. (2003) (in Chinese)[11]X.H. Zhu, G.W. Liu, and G.X. Ru, et al, “Community quantitative characteristics and dynamics of endangered plantspecies Abies chensiensis”, Chinese Journal of Ecology, 26(12), 1942-1946. (2007) (in Chinese)[12]X.M. Xiao, S. Boles, and J.Y.Liu, et al, “Characterization of forest types in Northeastern China, using multi-temporal SPOT-4 VEGETATION sensor data”, Remote Sensing of Environment, 82, 335–348(2002).[13]S. Gopal, C.E. Woodcock, and A.H. Strahler, “Fuzzy neural network classification of global land cover from1°AVHRR data set”, Remote Sens. Environ, 67,230-243 (1999).[14] C. O. Justice, & J. R. G. Townshend, “Special issue on the Moderate Resolution Imaging Spectroradiometer(MODIS): A new generation of land surface monitoring”. Remote Sensing of Environment, 83, 1−2 (2002).[15]X. Li, X. Cui, and X.D. Huang, et al, “Spatial and Temporal Change of MODIS vegetation indices for differentgrassland in northern Xingjiang”, Prataculrural Science, 24(9), 5-11 (2007). [in Chinese][16]X.F. Yu, and D.F. Zhuang, “Monitoring forest phenophases of Northeast China based on MODIS NDVI data”,Resources Science, 28(4), 111-117 (2006). [in Chinese][17]M.White, P. Thornton, & S. Running, “A continental phenology model for monitoring vegetation responses tointerannual climatic variability”. Global Biogeochemical Cycles, 11, 217-234 (1997).[18] B. Reed, J. Brown, D. Zee Vander, T. Loveland, et al., “Measured phenological variability from satellite imagery”.Journal of Vegetation Science, 5, 703-714 (1994).[19]N. Viovy, O. Arino, & A. S. Belward,, “The Best Index Slope Extraction (BISE): a method for reducing noise inNDVI time-series”. International Journal of Remote Sensing, 8, 1585-1590 (1992).[20]T. R. Loveland, B. C. Reed, J. F. Brown, et al. “Development of a global land cover characteristics database andIGBP DISCover from 1km AVHRR data”. International Journal of Remote Sensing,(1998).[21] B.K. Wyatt, J. N. Greatorex-Davies, M.O. Hill, et al, Comparison of Land Cover Definitions (Countryside 1990series, Vol.3). London: Department of the Environment (1994).[22]State Bureau of Surveying and Mapping, The technical regulation of global mapping, Beijing: State Bureau ofSurveying and Mapping, (1999).[in Chinese][23]J. E. Volgamann, T. Sohl, S. M. Howard. “Regional characterization of land cover using multiple sources of data”.Photogrammetric Engineering and Remote Sensing, 64(1):45~57(1998).[24]Yousong ZHAO, Anping LIAO, Lijun CHEN, “ The creation of national land cover database in China” Science ofSurveying and Mapping.32(4):138-140 (2007). [in Chinese][25]Pengfeng XIAO, Shunxi LIU, Xuezhi FENG, et al. “A Land Use/Cover Classification System Based on MediumResolution Remote Sensing Data”,China Land Science. 20(2),33-38(2006).[in Chinese][26]Zemeng FAN, Tianxiang YUE, Jiyuan LIU, et al. “Spatial and Temporal Distribution of Land Cover Scenarios inChina”. Acta Geographica Sinica ,60(6):941-952(2005). [in Chinese][27]R. G. Congalton, “A review of assessing the accuracy of classifications of remotely sensed data”. Remote Sens.Environ., 37, 35-46 (1991).[28]J.A. Richard, “Classifier performance and map accuracy”. Remote Sens. Environ., 57, 161-166 (1996).[29]S.V. Stehman, “Selecting and interpreting measures of thematic classification accuracy”, Remote Sens. Environ., 62,77-89(1997).[30]M. Story and R.G. Congalton, “Accuracy assessment: a user’s perspective”. Photogrammetric Engineering &Remote Sensing, 48(1), 131-137 (1986).。