| 摘要: |
| 为建立不依赖时序数据的水稻生育期识别模型，基于四波段辐射计(SKYE)获取的水稻全生育期每日的冠层光谱反射率数据，利用K近邻(k-nearest neighbors，KNN)、决策树(Decision trees)、支持向量机(Support vector machines，SVM)、随机森林(Random forests，RF)和梯度提升决策树(Gradient boosted decision trees，GBDT)共5种机器学习算法开展水稻生育期识别研究。结果表明:RF算法的识别准确率最高，达93.00%，KNN算法的识别准确率也达到了91.92%，其他3种算法的准确率也都超过90%。在此基础上，将建立的水稻生育期识别模型应用至无人机(UAV)影像数据，KNN算法适用性最好，识别准确率为83.54%，RF算法的适用性一般，识别准确率为74.38%，SVM算法的适用性最差，识别准确率仅为62.92%，但5种机器学习算法都容易错误地将抽穗扬花期识别为拔节孕穗期；而新构建的KNN算法结合可见光大气修正指数(Visible atmospherically resistant index，VARI)的水稻生育期识别模型对无人机数据的识别准确率可达86.04%，与单独应用KNN算法相比，对水稻各个生育期的识别精度更加均衡。 |
| 关键词: 水稻 生育期 光谱反射率 机器学习 植被指数 |
| DOI：10.11841/j.issn.1007-4333.2020.01.09 |
| 投稿时间：2019-01-11 |
| 基金项目:中央高校基本科研基金(2042017kf0236)；国家高技术研究发展计划(863计划)(2013AA102401) |
|
| Recognition of the rice growth stage by machine learning combined with vegetation index threshold |
|
YANG Zhenzhong,FANG Shenghui,PENG Yi,GONG Yan,WANG Dong
|
| (School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079; China) |
| Abstract: |
| In order to establish the rice growth period recognition model independent of time series data, a four-band radiometer(SKYE)was firstly used to collect the daily canopy spectral reflectance during the whole growth stage of rice in this study. Five common machine learning algorithms including K nearest neighbors(KNN), decision trees, support vector machine(SVM), random forests(RF)and gradient boosted decision trees(GBDT)were then used to recognize the rice growth stages. The results showed that the recognition accuracies of RF and KNN were 93. 00% and 91. 92%, respectively. The accuracy of the rest three algorithms were all over 90%. On this basis, when the rice growth stage recognition model was applied to aerial-collected data, KNN algorithm displayed the best applicability with the recognition accuracy of 83. 54%, SVM algorithm showed the worst applicability with the recognition accuracy of 62. 92%, and the recognition accuracy of RF algorithm reached 74. 38%. However, all the five algorithms mistakenly recognized the heading and flowering stages as jointing and booting stages. Here, a KNN algorithm combined with visible atmospherically resistant index(VARI)for rice growth stage recognition was proposed. The recognition accuracy of the model for aerial-collected data was up to 86. 04%, which was more balanced than that using KNN algorithm alone in recognizing different growth stages of rice. The results showed that KNN algorithm combined with VARI had higher accuracy and better generalization ability in recognizing rice growth period. |
| Key words: paddy rice growth stage spectral reflectance machine learning vegetation index |
