基于MODIS遥感数据的宏观土地覆盖特征分类方法与精度分析研究

针对宏观土地覆盖遥感分类的现状,充分利用MODIS相对于AVHRR数据具有的多光谱和分辨率优势,提出了利用MODIS数据进行分类特征选择与提取并结合多时相特征进行宏观土地覆盖分类的分类方法,并在中国山东省进行了分类试验,得出以下结论:①不同比例下的训练样本与验证样本影响着总体分类精度;②从MODIS数据中得到的植被指数EVI、白天地表温度Tday、水体指数NDWI、纹理特征局部平稳Homogeneity等可以作为分类特征配合参与到多波段地表反射率Ref1-7遥感影像中,能明显提高分类精度,而土壤亮度指数NDSI则没有贡献;③提取的分类特征对总体分类精度贡献大小为:EVI贡献最大,提高近6个百分点,其次是Homogeneity、NDWI,均提高近4个百分点,而最少的Tday也贡献了近3个百分点;④各分类特征对不同地物类别具有不同的分离度,在提高某些类别的分离性时,有可能降低了其它类别的分离性。试验结果表明:在没有其它非遥感信息的前提下,仅利用MODIS遥感自身信息对宏观土地覆盖分类就可达到较高精度。

Regional Land Cover Image Classification and Accuracy Evaluation Using MODIS Data

Based on the advantage of MODIS multi-spectrum data, this research explored a classificationmethod of feature selection and extraction, which combines the multi-spectrum data with multi-temporarydata in order to improve the classification accuracy. The classification accuracy was tested using 250 mMODIS data in Shandong province of China. The classification features were selected and extractedthrough the measures of the fractional cover, moisture, soil brightness, land surface temperature per day,and textures of various land cover types.The result indicates that it has higher classification accuracy using EVI (Enhanced Vegetation Index)as input than NDVI (Normalized Difference Vegetation Index), and NDWI (Normalized Difference WaterIndex) is superior to NDMI (Normalized Difference Moisture Index).The homogeneity of texture is thebest one for feature selection among the eight textures, and the optimal window size of texture is 11×11pixels. Whereas, NDSI (Normalized Difference Soil Index) almost has no effect for improving theclassification accuracy. For the contribution on improving the classification accuracy, EVI is the most, andthe following are homogeneity, NDWI and Tday(land surface temperature of day). The overall accuracyincreased about 10% through the method. The result shows that the feature selection and extraction canobviously improve classification accuracy, and the relatively high classification accuracy can also beacquired using the MODIS data sets without accessorial knowledge by this method.