华南地区典型种植园地遥感分类研究
作者简介:关舒婧(1993-),女,陕西华阴人,硕士生,主要从事遥感与GIS应用研究。E-mail: sj.guan@siat.ac.cn
收稿日期: 2017-06-19
要求修回日期: 2017-08-03
网络出版日期: 2017-11-10
基金资助
深圳市科技计划项目(JCYJ20150831194835299)
国家重点研发计划子课题(2016YFC0500201-07)
Study on the Classification of Typical Plantations in South China
Received date: 2017-06-19
Request revised date: 2017-08-03
Online published: 2017-11-10
Copyright
华南地区种植园地广泛分布,类型混杂多样,导致园地分布信息难以正确获取,为农业管理造成了较大困难。本研究基于Landsat8 OLI数据,通过数据融合、特征优化,应用随机森林算法构建面向对象的种植园地分类规则集,对华南地区典型经济作物香蕉、柑橘、葡萄、蒲葵、海枣、番木瓜和火龙果等进行类别识别,同时对比贝叶斯分类法、K最邻近分类法、支持向量机法、决策树分类法的分类效果。结果表明:数据融合会在一定程度上影响分类结果精度;植株形态、光谱特征接近,种植期交错是影响华南地区典型园地分类精度的重要原因;以中分辨率影像为数据源,面向对象的随机森林算法应用于种植园地分类研究总体精度可达88.05%,Kappa系数0.87,可以有效区分华南地区典型种植园地类别;相比于其他算法,随机森林算法在分类精度、可靠性和稳定性上具有一定优势,可为园地作物生长监测和种植管理提供科学依据。
关键词: 随机森林算法; 面向对象分类; 种植园地; landsat8 OLI; 数据融合
关舒婧 , 韩鹏鹏 , 王月如 , 韩宇 , 易琳 , 周廷刚 , 陈劲松 . 华南地区典型种植园地遥感分类研究[J]. 地球信息科学学报, 2017 , 19(11) : 1538 -1546 . DOI: 10.3724/SP.J.1047.2017.01538
Plantation refers to the land of perennial woody and herbaceous plants that are planted to collect fruit and leaves, including the land used for nursery. Plantations in southern China is widely distributed, and the types are mixed and diversified. As a result, it is difficult to obtain the information of the garden distribution, which has caused great difficulties for agricultural management. In order to improve the classification accuracy of the remote sensing images of the plant species, using Landsat8 OLI data, we took the random forest algorithm to construct object-oriented plantation classification rule set based on data fusion and feature optimization. This was used to classify the banana, citrus, grape, livistona chinensis, phoenix dactylifera, carica papaya, hylocereus undatus and so on, which are typically planted in southern China. At the same time, we compared the classification effects of Bayesian classification, K nearest neighbor classification, support vector machine method, decision tree classification method and random forest classification, and it verified the applicability of the object-oriented random forest approach for the classification of the garden types in medium-resolution images. Then, based on the clustering matrix of the classification results of random forest classification and the feature distance matrix between two categories, we analyzed the reasons that affect the classification accuracy of the garden by combining the image conditions and field survey results. Finally, we compared the characteristic difference of typically easy-to-mix land before and after data fusion, and evaluated the effect of data fusion on the classification of garden. The results show that: after the data fusion, the difference of the characteristics of water body and aquatic vegetation is reduced, and the spectral difference of livistona chinensis and phoenix dactylifera is reduced in some bands. Although, the data fusion can improve the image spatial resolution, to a certain extent, it weaken the spectral differences of objects. Thus, it would affect the classification effect based on spectral information. The plant morphology and spectral characteristics of the plantation land are close to each other, the planting period of the plantation land is intertwined and there are important factors that affect the classification accuracy of typical garden in southern China. Based on the medium resolution images, the object-oriented random forest algorithm for plantation classification can reach 88.05% and Kappa coefficient of 0.87, which can effectively distinguish the typical plantation land types in South China. The classification results of object-oriented stochastic forest algorithm show that the area of various plantations in the study area: banana covers an area of 700.2 hectares, citrus covers an area of 981 hectares, livistona chinensis covers an area of 81 hectares, phoenix dactylifera covers an area of 68.04 hectares, carica papaya covers an area of 93.24 hectares, hylocereus undatus covers an area of 167.4 hectares, and grape covers an area of 16.2 hectares. For the classification accuracy of the coverage types, the random forest classification was higher than that of the other four algorithms except grape and hylocereus undatus. The random forest algorithm has certain advantages in classification accuracy, reliability and stability, and can provide scientific basis for crop growth monitoring and planting management.
Fig. 1 Location of the study area图1 研究区位置示意图 |
Fig. 2 ESP calculation results图2 ESP计算结果 |
Fig. 3 Feature optimization results (raw data)图3 特征优化结果(原始数据) |
Tab. 1 Optimized selected features (raw data)表1 优化后所选特征(原始数据) |
指标 | 特征 |
---|---|
均值 | 相关性、信息熵、协同性、二阶距、第二、第三主成分、近红外波段、短波红外1、短波红外2 |
标准差 | 二阶距、协同性、相异性、相关性、信息熵 |
比率 | 短波红外1、短波红外2 |
对于邻域的平均差分 | 绿度分量、湿度分量 |
形状特征 | 长宽比 |
自定义 | 绿度分量、湿度分量、RVI、DVI |
Fig. 4 Flow chart of classification图4 分类流程图 |
Tab. 2 The total accuracy and kappa coefficients of the different classification methods表2 各组实验不同分类方法的总精度及kappa系数 |
方法 | 实验A | 实验B | |||
---|---|---|---|---|---|
总精度/% | kappa系数 | 总精度/% | kappa系数 | ||
Bayes | 53.22 | 0.45 | 58.73 | 0.47 | |
KNN | 79.73 | 0.77 | 77.59 | 0.74 | |
SVM | 80.30 | 0.77 | 79.89 | 0.77 | |
决策树 | 78.98 | 0.76 | 73.75 | 0.70 | |
随机森林 | 88.05 | 0.87 | 86.02 | 0.84 |
Fig. 5 Classification rule set of experiment A图5 实验A分类规则集 |
Fig. 6 Producer accuracy and user accuracy of five classification methods of experiment A图6 实验A 5种分类法生产者精度及用户精度比较 |
Fig. 7 Spatial distribution of random forest, decision tree, SVM, KNN, Bayes图7 随机森林、决策树、SVM、KNN、Bayes分类空间分布图 |
Tab. 3 Category classification matrix表3 类别分类矩阵 |
水库/坑塘 | 番木瓜 | 葡萄 | 香蕉 | 其他 | 火龙果 | 裸土 | 柑橘 | 蒲葵 | 海枣 | 水生植被 | 建设用地 | |
---|---|---|---|---|---|---|---|---|---|---|---|---|
水库/坑塘 | 0.0 | 9.2 | 9.8 | 8.6 | 7.7 | 10.1 | 7.6 | 6.6 | 7.4 | 8.5 | 3.1 | 20.3 |
番木瓜 | 9.2 | 0.0 | 4.2 | 2.2 | 3.3 | 2.2 | 6.0 | 2.1 | 2.3 | 2.6 | 3.9 | 12.7 |
葡萄 | 9.8 | 4.2 | 0.0 | 9.8 | 6.8 | 3.9 | 3.1 | 2.6 | 6.2 | 7.7 | 5.4 | 6.2 |
香蕉 | 8.6 | 2.2 | 9.8 | 0.0 | 2.4 | 4.0 | 10.2 | 2.7 | 3.4 | 3.9 | 4.4 | 20.7 |
其他 | 7.7 | 3.3 | 6.8 | 2.4 | 0.0 | 3.2 | 7.9 | 2.5 | 4.9 | 4.8 | 3.8 | 15.6 |
火龙果 | 10.1 | 2.2 | 3.9 | 4.0 | 3.2 | 0.0 | 8.8 | 3.4 | 4.3 | 5.9 | 4.0 | 14.8 |
裸土 | 7.6 | 6.0 | 3.1 | 10.2 | 7.9 | 8.8 | 0.0 | 2.4 | 7.0 | 5.2 | 5.8 | 5.4 |
柑橘 | 6.6 | 2.1 | 2.6 | 2.7 | 2.5 | 3.4 | 2.4 | 0.0 | 3.6 | 2.3 | 2.8 | 8.0 |
蒲葵 | 7.4 | 2.3 | 6.2 | 3.4 | 4.9 | 4.3 | 7.0 | 3.6 | 0.0 | 2.3 | 3.6 | 17.9 |
海枣 | 8.5 | 2.6 | 7.7 | 3.9 | 4.8 | 5.9 | 5.2 | 2.3 | 2.3 | 0.0 | 3.6 | 15.4 |
水生植被 | 3.1 | 3.9 | 5.4 | 4.4 | 3.8 | 4.0 | 5.8 | 2.8 | 3.6 | 3.6 | 0.0 | 15.5 |
建设用地 | 20.3 | 12.7 | 6.2 | 20.7 | 15.6 | 14.8 | 5.4 | 8.0 | 17.9 | 15.4 | 15.5 | 0.0 |
Tab. 4 Error matrix of random forest classification results表4 随机森林分类结果的误差矩阵 |
用户类别 | 参考类别 | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|
番木瓜 | 葡萄 | 香蕉 | 水库/坑塘 | 其他 | 火龙果 | 裸土 | 柑橘 | 蒲葵 | 海枣 | 水生植物 | 建设用地 | 合计 | |
番木瓜 | 14 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 14 |
葡萄 | 0 | 5 | 0 | 0 | 0 | 0 | 2 | 0 | 0 | 0 | 0 | 0 | 7 |
香蕉 | 0 | 0 | 85 | 0 | 2 | 0 | 0 | 3 | 0 | 0 | 1 | 0 | 91 |
水库/坑塘 | 0 | 0 | 0 | 100 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 0 | 104 |
其他 | 0 | 0 | 0 | 0 | 48 | 0 | 0 | 0 | 1 | 0 | 1 | 0 | 50 |
火龙果 | 0 | 0 | 0 | 0 | 1 | 6 | 0 | 0 | 0 | 0 | 0 | 0 | 7 |
裸土 | 0 | 0 | 0 | 0 | 0 | 0 | 76 | 0 | 0 | 0 | 0 | 0 | 76 |
柑橘 | 1 | 2 | 0 | 0 | 2 | 0 | 5 | 64 | 1 | 0 | 2 | 1 | 78 |
蒲葵 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 5 | 2 | 0 | 0 | 7 |
海枣 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 12 | 0 | 0 | 12 |
水生植物 | 0 | 0 | 0 | 2 | 3 | 0 | 0 | 0 | 0 | 0 | 53 | 0 | 58 |
建设用地 | 0 | 0 | 0 | 0 | 1 | 0 | 5 | 0 | 0 | 0 | 0 | 18 | 24 |
合计 | 15 | 7 | 85 | 102 | 57 | 6 | 88 | 67 | 7 | 14 | 61 | 19 |
Tab. 5 The easy mix of land and analysis of its causes表5 地类易混分组合及其原因分析 |
地类易混分组合 | 原因分析 |
---|---|
水库/坑塘与水生植被 | 在影像获取的时间,研究区的水生植被藕塘已处于枯萎阶段,水体露出,故存在混分性 |
建设用地和裸土 | 研究区内裸土为休耕期土地,与待建的建设用地在光谱、纹理上确实极为接近 |
番木瓜与柑橘 | 番木瓜和柑橘在种植初期,植株小,植被覆盖度低,土壤信息突出,在成熟期光谱特征也较为接近,故易混分 |
葡萄与柑橘 | 葡萄属藤本植物,在该时像下植被盖度、绿度均不高,故易与种植初期的柑橘混分 |
蒲葵与海枣 | 除蒲葵呈灰绿色与海枣存在差异外,蒲葵与海枣植株形态接近,生长阶段接近,易混分性强 |
火龙果与番木瓜 | 火龙果种植间隙大,在30 m分辨率的影像上除植被信息外还有土壤信息混在其中,降低了光谱特点,故易与番木瓜存在混分 |
Fig. 8 Comparison of the characteristics difference before and after Data Fusion图8 数据融合前后地物特征差异对比 |
The authors have declared that no competing interests exist.
[1] |
[
|
[2] |
[
|
[3] |
[
|
[4] |
|
[5] |
[
|
[6] |
[
|
[7] |
[ Wang Y, Fu M C, Wang L, et al. Tree-cotton intercropping land extraction based on multi-source high resolution satellite imagery[J]. Remote Sensing for Land & Resources, 2017,29(2):152-159. ]
|
[8] |
[
|
[9] |
[
|
[10] |
[
|
[11] |
[
|
[12] |
[
|
[13] |
[
|
[14] |
|
[15] |
[
|
[16] |
[
|
[17] |
[
|
[18] |
[
|
[19] |
[
|
[20] |
[
|
[21] |
|
[22] |
[
|
/
〈 | 〉 |