基于随机森林和纹理特征的苹果园遥感提取

Apple orchard remote sensing extraction based on random forest and texture features

为准确、快速地从高分影像中获取苹果种植分布信息,以QuickBird遥感影像为数据源,首先采用分形理论和灰度共生矩阵(GLCM)提取纹理特征,然后将提取的分形纹理和GLCM纹理特征分别与光谱特征组合,最后开展随机森林分类实验,对不同分类特征和不同分类方法的实验结果进行比较。结果表明:光谱+GLCM纹理识别苹果园的效果明显优于光谱特征和光谱+分形纹理,其苹果园提取精度为95.99%,比光谱分类显著提高11.83%,比光谱+分形纹理提高1.34%;在相同分类特征下随机森林分类结果最高,其中,随机森林结合光谱+GLCM纹理分类精度最高,总体精度和Kappa系数分别为95.30%和0.94,较最小距离和支持向量机分类有明显提高。

In order to obtain the information of apple cultivation and distribution by high resolution image accurately and rapidly,QuickBird remote sensing image is taken as the data source,and fractal theory and gray level co⁃occurrence matrix(GLCM)are adopted to extract the texture features,then the extracted fractal texture features and GLCM texture features are combined with spectrum features respectively.The random forest classification experiments are conducted,and the experimental results of different classification features and classification methods are compared.The results show that effect of spectrum plus GLCM texture is obviously superior to that of spectrum features and spectrum plus fractal texture in identifying apple orchard.This method′s extraction accuracy of apple orchard is 95.99%,which is significantly 11.83%higher than that of spectrum classification and 1.34%higher than that of spectrum plus fractal texture.The effect of the random forest classification is the best under the condition of the same classification features,among which random forest classification combining spectrum plus GLCM texture classification achieved the highest accuracy(overall accuracy and Kappa coefficient are 95.30%and 0.94 respectively),which has been improved significantly in comparison with minimum distance classification and support vector machine(SVM)classification.