| 摘要: |
| 【目的】准确的作物空间分布是农业估产、作物长势和病虫害防控等农业遥感监测的重
要基础信息。选择合适的特征和分类器对作物空间信息的提取有重要意义。【方法】文章基于
北安市的Landsat 8时间序列数据探究了特征提取和分类器选择对作物分类的影响。首先，基
于Google Earth Engine（GEE） 平台提取光谱、植被指数、纹理和物候时序特征；其次，将不
同特征及其组合输入最小距离法（Minimum Distance Classification，MDC）、朴素贝叶斯
（Na?ve Bayes，NB）、K最近邻法（K-Nearest Neighbor，KNN）、支持向量机（Support Vector
Machine，SVM） 和随机森林（Random Forest，RF） 5种分类器比较精度；最后，计算分离性
指数（Separability Index，SI） 评估特征对识别作物的贡献度，辅助验证分类器的分类结
果。【结果】研究结果表明：（1） 4类特征中光谱特征分类精度最高，3种特征组合中光谱+植
被指数精度最高，但相较于光谱特征精度提仅提高0.6%，说明时序光谱特征足以得到较好的
作物分类结果，提取的其他特征对精度提升作用不明显；（2） 通过比较5种分类器的精度均值
和标准差，性能最好的是RF，其次是SVM，MDC的性能最差；（3） 在特征分离性方面，光谱
特征最好，其次是植被指数、物候和纹理特征。【结论】光谱时序特征结合RF分类器效率最
高，能得到较好的作物识别效果。文章能为作物分类特征提取和分类器选择提供参考和依据。 |
| 关键词: 遥感 时间序列 作物分类 特征优选 分类器 |
| DOI：10.12105/j.issn.1672-0423.20210101 |
| 分类号: |
| 基金项目:科技基础资源调查专项（No.2019FY202501）；国家自然科学基金项目“耕地规模化利用的多尺度智能遥感
监测方法研究”（42071419）；中央级公益性科研院所基本科研业务费专项“地块尺度的耕地利用规模时空
变化研究”（1610132020016）；气候变化对中国小麦主产区生产系统影响综合评价（2017YFD0300201-3） |
|
| A comparative analysis of feature extraction and classifiers forcrop classification based on time series data |
|
Wen Caiyun, Lu Miao※, Song Qian, Cheng Rui, Zhang Shibo
|
|
Key Laboratory of Agricultural Remote Sensing，Ministry of Agriculture and Rural Affairs / Institute of Agricultural
Resources and Regional Planning，Chinese Academy of Agricultural Sciences，Beijing 100081，China
|
| Abstract: |
| ［Purpose］Accurate spatial distribution of crops is important basic information for
agricultural remote sensing monitoring such as yield estimation， crop growth and pest control.
［Method］In this study，we designed a series of experiments to evaluate what features are most
useful and how various classifiers affect the crop-type classification performance. Firstly， we
extracted the spectral，vegetation index，textural and phenology features based on the Landsat 8
time series data from Google Earth Engine （GEE） in Bei’an city，and then utilized Minimum
Distance Classification （MDC），Na?ve Bayes （NB），K-Nearest Neighbor （KNN），Support
Vector Machine （SVM），and Random Forest （RF） classifiers for crop classification based on
the above features and their combinations. Lastly，the separateness index （SI） was employed to
measure the feature contribution to crop identification and assist in validating the classification
results.［ Result ］ The results show that the spectral features have the highest classification
accuracy among the four types of features，and the combination of spectral and vegetation index
features improves only 0.6% over the spectral features. Therefore， it is reasonable to consider
that the feature extraction does not play a significant role in improving classification accuracy. By comparing the mean and standard deviation of accuracies，the best performance among the five
classifiers is RF，followed by SVM and the worst is MDC. For feature importance，the spectral
features play the most important role for crop classification，the followings are vegetation index，
phenology and texture features.［Conclusion］Spectral features combined with RF is the most
efficient combination. |
| Key words: remote sensing time series crop classification feature optimization classifier |
