摘要:准确获取果树的空间种植分布信息，对于开展果树长势监测、产量估算等具有重要意义。为提取太湖流域金庭镇果树的空间分布，本研究以冬夏时期的两景高分二号（GF-2）遥感影像为数据源，利用归一化植被指数（NDVI）和归一化水体指数（NDWI）结合纹理特征构建了基于光谱指数和纹理特征的决策树模型，提取了金庭镇2017年果树的空间分布信息。通过分析研究区各地类的光谱曲线发现，植被与非植被区分明显，但果树与茶树的光谱存在混淆。GF-2影像包含丰富的纹理信息，果树与茶树在GF-2影像上纹理特征明显，易于区分。纹理可作为果树提取的重要特征。为了确定最佳纹理窗口的大小，研究中提出了累计差（Δf）的方法。通过比较每一个纹理变量在15种不同尺度窗口（3×3，5×5，7×7，9×9，11×11，13×13，15×15，17×17，19×19，21×21，23×23，25×25，27×27，29×29，31×31）下的Δf，确定了最佳纹理窗口为15×15。在最佳纹理窗口下根据累计差选取了5大纹理组合：均值（mean）、方差（variance）、对比度（contrast）、信息熵（entropy）和相关性（correlation）。研究结果表明基于光谱指数NDVI和NDWI结合纹理特征构建的决策树模型可有效区分果树与茶树。累计差的方法能够快速确定最佳纹理窗口和纹理组合。提取结果说明果树分布于金庭镇的各个位置，主要分布在平原区，种植比较整齐，南部种植面积多于北部。本研究果树的提取精度为95.23%，模型总体分类精度为89.57%，Kappa系数为89.00%，果树的生产精度为90.00%，用户精度为87.30%。与单一光谱、纹理模型相比，本文模型总体分类精度更高，精度分别提升了10.65%和12.04%。该方法能够适用于大区域果树的遥感提取，可为亚米级遥感影像研究果树的纹理特征提供重要参考和借鉴价值。此外，文中提出的累计差可为选取最佳纹理窗口提供一种新的思路。
关键词:GF-2影像/
果树提取/
光谱特征/
纹理特征/
决策树分类
Abstract:The accurate acquisition of planting area and spatial distribution information is essential to monitor the growth and estimate the production of fruit trees (orchard). Remote sensing has been widely used in crop identification and monitoring in recent decades. Numerous classification algorithms have been developed based on various requirements for remote sensing data analysis. However, distinguishing fruit tree orchard and tea garden remains challenging, due to their similar spectral characteristics. Two GF-2 WFV (wide field of view) images, taken in summer and winter, were used to extract the spatial distribution of fruit trees in Jinting Town in the Taihu Lake Basin in this study. The normalized difference vegetation index (NDVI), normalized difference water index (NDWI), and texture features were used to construct a decision tree model. Vegetation and non-vegetation were quickly identified by analyzing the spectral curves of ground features in the study area. However, spectral characteristic was a poor parameter to differentiate fruit trees from tea trees. Since fruit trees and tea trees have distinct textural features, GF-2 images with rich texture information on ground objects can help distinguish fruit trees from tea trees. Thus, texture is one of the most important features in fruit tree extraction. In this study, the method of cumulative difference (Δf) was used to determine the optimal size of the texture window. Among the Δf values of each texture under 15 different window scales (3×3, 5×5, 7×7, 9×9, 11×11, 13×13, 15×15, 17×17, 19×19, 21×21, 23×23, 25×25, 27×27, 29×29, 31×31), the 15×15 window was determined as the optimum texture window. In addition, five texture features that were easy to distinguish from other objects were selected according to the cumulative difference of variables such as mean, variance, contrast, entropy, and correlation under the optimal texture window. The results showed that the decision tree model based on spectral index NDVI and NDWI, combined with texture features, effectively distinguished fruit trees from tea trees. The method of cumulative difference can quickly determine the best texture window size and texture combination. The extraction results showed that fruit trees were widely distributed in all locations of Jinting Town and that the planting area in the south was larger than that in the north. The local detail map indicated that the distribution of fruit trees was relatively neat and mainly in the plain area. The extraction accuracy of fruit trees in this study was 95.23%. The overall accuracy of the model in this study was 89.57% and the kappa coefficient was 89.00%. The producer accuracy and user accuracy were 90.00% and 87.30%, respectively. Using spectral indices combined with textural features achieved a higher overall accuracy than using spectral indices or textural features alone, with an overall accuracy increase of 10.65% and 12.04%, respectively. This method can be applied to the remote sensing extraction of fruit trees on a large scale and can provide an important reference in fruit tree extraction by using texture characteristics of sub-meter images. Moreover, the cumulative difference proposed in this study provides a new method for selecting the best texture window.
Key words:GF-2 images/
Fruit tree extraction/
Spectral features/
Textural features/
Decision tree classification
PDF全文下载地址:
http://www.ecoagri.ac.cn/article/exportPdf?id=2e199476-7749-4d9c-8e84-a0ef548b5f33
