| 刘杰,刘吉凯,安晶晶,章超.基于时序Landsat 8 OLI多特征与随机森林算法的作物精细分类研究[J].干旱地区农业研究,2020,38(3):281~288 |
| 基于时序Landsat 8 OLI多特征与随机森林算法的作物精细分类研究 |
| Precise crop classification based on multi\|features from time\|series Landsat 8 OLI images and Random Forest Algorithm |
| |
| DOI：10.7606/j.issn81000-7601.2020.03.37 |
| 中文关键词: 随机森林算法 作物分类 时序Landsat 8 OLI 特征重要性 新疆 |
| 英文关键词:Random Forest Algorithm crops classification time\|series Landsat 8 OLI images variable importance Xinjiang |
| 基金项目:淮河流域气象开放研究基金（HRM201606）;安徽科技学院引进人才资助项目（ZHYJ201603） |
|
| 摘要点击次数: 772 |
| 全文下载次数: 742 |
| 中文摘要: |
| 利用新疆阿克苏地区温宿县2014—2015年生长季的7景时序Landsat 8 OLI数据，提取光谱特征、纹理特征、植被指数等高维信息，并基于随机森林算法（Random Forest, RF）构建分类模型。分析了RF模型中重要参数树个数k、节点分裂特征个数m对分类精度的影响，计算GINI系数评估所有特征重要性，探索最佳特征子集，完成模型的参数率定与信息冗余消除，实现了温宿县研究区内的多种作物类型精细分类，并对比分析了随机森林与其他几种机器学习算法的分类性能。结果表明：作物分类的3类特征中，重要性排名靠前的分别是影像纹理平均规则程度Mean、与作物水分含量密切相关的地表水分指数（LSWI）及短波红外波段光谱反射率，对应干旱区作物的2个关键时相：生长旺盛期与播种期；随机森林分类精度受分类特征数量的影响。当特征删除量低于总特征数的30%时，RF模型的分类精度基本保持不变；当删除量超过70%时，分类精度下降的幅度加大；随机森林方法相对于决策树、支持向量机、朴素贝叶斯、K-近邻等监督分类算法，无论是分类结果的精度还是分类效率均具有优势。 |
| 英文摘要: |
| According to the 7-scene landsat 8 OLI data of the 2014-2015 growth season in Wensu County, Aksu region, Xinjiang, we extracted spectral features, texture features, vegetation index, and other high\|dimensional information, and constructed a classification model based on the Random Forest (RF) Algorithm. The influence of the number of important parameter trees k and the number of node splitting features m in the RF model on classification accuracy was analyzed, and GINI coefficient was calculated to evaluate the importance of all features to explore the best feature subset. In this study, parameter calibration and information redundancy elimination of the model were completed, and the precise classification of various crop types from the study area of Wensu County was realized. The classification performance of RF and other machine learning algorithms was compared and analyzed. The results showed that among the three characteristics of crop classification, the most important ones were mean of image texture, land surface water index (LSWI) closely related to crop moisture content, and spectral reflectance of short\|wave infrared, which were corresponding to two key time phases of crops in arid areas: growth period and sowing period. Moreover, the classification accuracy of RF was affected by the number of classification features. When the deletion amount of “feature” was less than 30% of the total feature number, the classification accuracy of RF model remains unchanged. When the deletion amount was more than 70%, the classification accuracy decline increased. Finally, compared with decision tree, support vector machine, naive Bayes, k-nearest neighbor, and other supervised classification algorithms, Random Forest Algorithm had advantages in both accuracy and efficiency of classification results. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| | |