基于纹理特征和SVM的QuickBird影像苹果园提取

为提高高空间分辨率遥感影像(高分影像)中苹果园提取精度,基于Quick Bird遥感数据,研究综合光谱特征和纹理特征的苹果园自动提取方法。该方法首先采用最佳指数因子(OIF)获取多光谱波段最佳组合,然后采用不同大小滑动窗口(从3像素×3像素到13像素×13像素)提取全色波段的灰度共生矩阵(GLCM)、分形和空间自相关3种纹理特征并分别与光谱特征组合,最后通过支持向量机(SVM)分类进行苹果园分类识别。研究表明:在分类特征上,与单一光谱或纹理特征相比,光谱特征结合纹理特征能有效提高苹果园提取精度(Fa)和总体分类精度(OA),其中光谱+GLCM纹理(9像素×9像素)分类精度最高,Fa和OA分别为96.99%和96.16%,比光谱+分形纹理分别提高0.63个百分点和1.56个百分点,比光谱+空间自相关纹理显著提高11.92个百分点和9.20个百分点。在分类方法上,通过对比分析SVM、最大似然和神经网络3种方法的分类结果,探明SVM分类识别苹果园精度最高。最后对苹果园提取结果进行面积统计,结果表明GLCM纹理结合SVM分类的苹果园面积估算与目视解译结果的一致性超过98%。

Apple Orchard Extraction with QuickBird Imagery Based on Texture Features and Support Vector Machine

In order to improve the accuracy of apple orchard extracting in very high spatial resolution (VHSR) remote sensing image, an automated apple orchard extracting method based on texture features together with spectral values and support vector machine (SVM) was studied. This method firstly obtained the optimum combination of multi-spectral bands by using the optimum index factor (OIF);then three kinds of texture features, namely gray level co-occurrence matrix (GLCM), fractal and spatial autocorrelation texture with six different window sizes (from 3 pixels×3 pixels to 13 pixels×13 pixels) were extracted from the panchromatic image for comparison, and further merged with spectral values respectively;finally the above features were used to identify apple orchard by using SVM. Experiments using QuickBird data showed that spectral features combined with texture features could achieve higher apple orchard extraction accuracy (Fa) and overall accuracy (OA) than using spectral features or textures features alone. Among the different features used, the spectral+GLCM features (with 9 pixels×9 pixels) achieved the highest accuracy (Fa and OA were 96.99% and 96.16%, respectively), which were slightly higher (0.63 and 1.56 percentages, respectively) than those of spectral+fractal features and significantly higher (11.92 and 9.20 percentages, respectively) than those of spectral+spatial autocorrelation features. Among the different classification methods, three classification techniques (SVM, maximum likelihood and neural networks) were compared for accuracy in apple orchard detection, and results suggested that SVM had the highest accuracy in identifying apple orchard. McNemar test was also computed for statistic significance among spectral+GLCM and other features and also among the three classifiers, and the confidence levels were all less than 5%. Consistency of the extracted apple orchard area and the visual interpretation results according to filed investigation and Google Earth VHSR concurrent image were able to achieve 98% in test regions.