结合纹理和空间特征的多光谱影像面向对象茶园提取

云南茶园主要分布于山区，往往与其他地物混合，破碎化程度高，给基于遥感的高精度茶园提取带来困难。破碎化山地茶园的遥感识别能力有待进一步提高。以西双版纳北部及普洱市南部山区为研究区，以高分一号(GF-1)遥感影像为数据源，基于易康(eCognition9.0)软件，采用多尺度分割(MRS)方法对影像进行分割，并通过ED3_Modified评估影像的最优分割尺度。首先构造了包括14个光谱特征、6个纹理特征和3个空间特征的23维初始特征，通过计算样本的分离度对分类特征空间进行优化，确定了16维最优特征空间。然后应用面向对象的监督分类方法贝叶斯(Bayes)、决策树5.0(DT5.0)及随机森林(RF)]对研究区茶园进行提取。采用实地调查样点和随机点结合的方法对提取结果进行精度验证，并比较了不同分类方法的茶园提取精度。面向对象的监督分类多分类（茶园、森林、农田、不透水层、水体）]结果的总体精度(OA)和Kappa系数分别为(Bayes: 87.73%, 0.70)，(DT: 88.52%, 0.72)，(RF: 91.23%, 0.78)。三种分类方法对茶园提取的生产者精度(PA)和使用者精度(UA)分别为(Bayes: 67.23%, 75.33%)，(DT: 68.84%, 83.83%)，(RF: 70.54%, 87.13%)；相比于面向对象的RF多分类，面向对象RF二分类（茶园、其他地物）OA和Kappa系数分别提高了3.24%和0.07，茶园提取的PA与UA分别提高了5.99%和5.61%；相较于仅利用光谱特征的基于像元的RF多分类，面向对象的RF二分类OA与Kappa系数分别提高了23.32%和0.27，茶园提取PA与UA分别提高了21.10%和29.03%。结果表明：采用面向对象的监督分类方法在对茶园提取方面有应用潜力，尤其面向对象的RF分类精度更高，二分类相较于多分类对于茶园提取更为精细和准确。该方法对于复杂、破碎山地茶园提取精度较高，能够满足基于高空间分辨率多光谱影像的茶园精细化识别应用需求。

Combining Textures and Spatial Features to Extract Tea Plantations Based on Object-Oriented Method by Using Multispectral Image

The tea plantations of Yunnan province are mainly fragmentally distributed in mountainous areas and often mixed with other ground objects, making it difficult to extract tea plantations with high precision based on remote sensing. Combining textures and spatial features based on Object-Oriented method is rarely applied to extract tea plantations in previous crop classification research using multi-spectral imagery. Therefore, it is necessary to explore further the recognition ability for tea plantations by using high spatial resolution and multi-spectral images under the fragmental and mountainous region. In this research, a typical mountainous area located between the northern Xishuangbanna autonomous prefecture and the southern of Pu’er city was used as our study area. A scene image with a 2 m resolution pan-chromatic and 8 m resolution multi-spectral derived from the GF-1 PMS sensor was used as the source data for our research. The eCognition Developer9.0 software was employed to segment the image by multi-resolution segmentation, and the ED3_Modified method was used to evaluate the optimal segmentation scale. Firstly，we constructed 23 dimensions of original features including 14 spectral features, 6 textures and 3 spatial features. Secondly, 16 optimal features were selected to classify by calculating the separation distance of five land-cover types. Thirdly, based on 16 optimal features space, three object-oriented supervised classification methods (Bayes, Decision Tree 5.0 (DT) and Random Forest (RF) were applied to extract the tea plantations of the study area. Finally, filed survey samples and random samples were used to validate the accuracy of tea plantations extraction results, and we compared the classification accuracies of different classification methods. The results showed that for the multi-classification (including tea plantations, forest, cropland, impervious and water body) the overall accuracy (OA)/ Kappa coefficient (Kappa) are Bayes (87.73%/0.70), DT5.0 (91.23%/0.78) and RF (88.52%/0.72) respectively, but for tea plantations, the producer accuracy (PA)/user accuracy (UA) are Bayes (67.23%/75.33%), DT (68.84%/83.83%) and RF (70.54%/87.13%). Compared with the object-oriented RF multi-classification, the OA and Kappa of the object-oriented RF binary classification (tea plantations and others) increased by 3.24% and 0.07, the PA/UA of tea plantations increased by 5.99%/5.61%. Similarly, compared with the pixel-based multi-classification, the OA and Kappa of the object-oriented RF binary classification increased by 23.32% and 0.27, the PA/UA of tea plantations increased by 21.10%/29.03%, respectively. The results indicated that the object-oriented supervised classification methods have the potential for tea plantations extraction, especially the object-oriented RF classification got a higher accuracy. Moreover, the binary classification method has higher accuracy than that of multi-classification for tea plantation extraction. Our object-oriented method that combined textures and spatial features with spectral features is effective for tea plantations extraction, especially when applied to the complex and fragmental mountainous landscape. Our method can meet the application requirements in fine tea plantations identification based on high-spatial resolution and multi-spectral imagery too.