| 摘要: |
| 【目的】遥感影像监督分类能够快速获取土地利用和地表覆盖的信息，分类样本的选
取对分类精度具有决定性的作用。以最大似然分类方法为例，研究样本数量、均值和标准差
对分类精度的影响。【方法】基于地表覆盖产品GlobeLand30 分层随机选取不同数量的训练
样本，采用最大似然法对研究区域的Landsat8 遥感影像进行分类。通过谷歌地球高分影像选
取一定数量的检验样本，对影像分类结果进行精度评价，并研究样本数量、均值和标准差对
分类结果的影响。【结果】不同数量的训练样本得到的分类精度不同，分类精度随着样本数
量的增加先增加后下降，然后渐趋于稳定；在样本质量特征方面，当训练样本的均值和标准
差越接近检验样本的均值和标准差时，分类结果的精度越高，反之则分类精度较低。【结论】
在最大似然分类过程中，训练样本数量的选取存在临界值，当达到临界值时即可获得较高分
类精度，随后即使增加样本的数量也无法显著提高分类结果的精度。在样本质量方面，要尽
量选取能够反映地物真实特征的训练样本进行分类。 |
| 关键词: 最大似然分类 样本数量 样本均值 样本标准差 分类精度 |
| DOI：10.12105/j.issn.1672-0423.20180207 |
| 分类号: |
| 基金项目:中央级公益性科研院所基本科研业务费专项（1610132018017） |
|
| Study on training sample sensitivity of maximum likelihood classification |
|
Lu Xiaoguo1,2, Wang Tongke1, Liang Shefang2, Lu Miao※2
|
|
1.School of Mathematical Sciences，Tianjin Normal University，Tianjin 300387，China;2.Key Laboratory of
Agricultural Remote Sensing，Ministry of Agriculture/Institute of Agricultural Resources and Regional Planning，
Chinese Academy of Agricultural Sciences，Beijing 100081，China
|
| Abstract: |
| ［Purpose］The surveillance classification of remote sensing images can quickly
obtain the information of land use and land cover，and the selection of classification samples plays
a decisive role in classification accuracy. This study takes maximum likelihood classification as an
example to study the effects of sample quantities，mean and standard deviation on classification
accuracy.［ Method］This study selects the samples of different quantities randomly based on
GlobeLand30，and applies the maximum likelihood method to classify Landsat8 images of study
area. After selecting a certain number of test samples from Google Earth，this study evaluate the
accuracy of classification results as well as research the effects of sample quantities，mean and
standard deviation on classification results.［ Result］The classification accuracies with different
training samples are different. With the increase of sample quantities，the classification accuracy
increases firstly and gradually becomes stable after a slight decrease. With respect to sample
quality characteristics，the accuracy of classification results is higher when the mean and standard
deviation of training samples are close to test sample. Otherwise，the classification accuracy is lower.
［Conclusion］There is a critical value in the selection of sample quantity during the process of the
maximum likelihood classification. When it reaches the critical value，the accuracy of the classification
result can be high. Subsequently，even if the number of samples is increased，the classification
accuracy cannot be significantly improved. In terms of sample quality，we should try to select training
samples that can reflect the real characteristics of ground objects. |
| Key words: maximum likelihood classification sample quantities sample mean sample
standard deviation classification accuracy |
