基于决策树和MODIS数据的土地利用分类

Land Use Classification based on Decision Tree Using MODIS Data

土地利用类型的识别是土地利用/土地覆盖研究中的重点内容,如何准确、快速的获得大尺度范围的土地覆盖信息进行土地变化的动态实时监测一直是关注的重点。本文针对MODIS数据多光谱的特点,以山东省为例,选取8月份8-day的一期MODIS09Q1、MODIS09A1产品及全年16-day的MODIS13Q1NDVI时间序列产品,通过分析各种土地利用类型的光谱间关系,同时选择NDVI、EVI、NDWI、NDMI、NDSI等分类指数,并构造新的波段B2/B1、B7/B6(B1、B2、B6、B7分别代表1波段、2波段、6波段、7波段),利用决策树分类法,进行土地利用分类试验。结果表明,仅利用MODIS数据自身信息对宏观的土地利用分类就可以达到较高的精度,分布范围完整的土地利用类型如耕地、城市居民点精度较高,零星分布的土地利用类型如农村居民点、草地分类精度较低。决策树分类法充分发挥了MODIS数据的多光谱特点,总体精度达到71.4%,kappa系数为0.68。相对于最大似然法,总体精度提高近10个百分点,对耕地及沼泽等类型的精度提高20%到25%。

Land use type recognition plays an important role in land use/cover research. People pay close attention to how to get the large-scale land cover information to monitor the land use changing. Compared with AVHRR data, MODIS data is better in spectrum, spatial resolution, data quality, and MODIS sensor has seven wave bands that are designed for the terrestrial system; so it attracts much attention in the mapping and monitoring the land use and land cover change in global or regional scale. Based on the advantage of MODIS multi-spectrum data, we choose MODIS products MODISO9Q1, MODISO9A1 for one period of 8-day and multi-temporal MODIS13Q1 NDVI data for a year of 16-day. From analyzing the relations between the spectrums, we select the classification features such as NDVI（Normalized Difference Vegetation Index）, EVI （ Enhanced Vegetation Index）, NDWI （ Normalized Difference Water Index）, NDMI （Normalized Difference Moisture Index）,NDSI（Normalized Difference Soil Index）, and make new bands such as B2/B 1, B7/B6（ B 1, B2, B6, B7 represent band 1, band2, band6, band7）, the purpose is to extract the land use information of Shandong province in China by the decision trees. The result indicates that The decision tree classifier can make classification processing simple and avoid the probability of confusion, the relatively high classification accuracy can be acquired using MODIS data without other assistant information, classification accuracy of large-scale objects is better than the scattered objects. Advanced technology selection and classification characters extraction can improve the accuracy too. Multi-spectrum is an important characteristic for MODIS data; Decision Tree classifier can take this advantage perfectly. Its overall classification accuracy is 71.4%, kappa is 0.68, compared to MLC （maximum likelihood classifier）, the overall accuracy can increase by ten percent, the accuracy of field and swamp can increase by 20 % ~ 25 %. The result shows that defining the threshold value of different land use type is the key in decision tree, One temporary data can＇t distinguish forest, grass and field. But different vegetation has different growth role. We can use the multi- temporary data such as the vegetation index to extract forest,grass and field. Otherwise, River＇s spectrum is especial. The deep water has typical water character, but the adlittoral water＇ spectrum feature is similar to country. Of course, this classification system is suit for the experiment area, whether can be put in use in other areas, we must research this continually in the future.