机器学习法的干旱区典型农作物分类

当前，基于机器学习方法开展农作物分类研究，对于确保干旱区粮食安全和生态安全有着极为重要的现实意义。基于机器学习方法，采用时间序列Sentinel 2A遥感数据提取农作物分类信息，通过引入地块基元和红边特征，探讨了不同分类特征组合对机器学习分类精度的影响。结果表明：随机森林分类器可以有效集成光谱和植被指数等多维向量的优势，将其应用于干旱区典型农作物分类上的精度均在89%以上，分类组总体精度最高可达94.02%。地块基元点集支持下的分类特征提取方法能够提高机器学习效率和农作物分类精度，使光谱组及指数组的分类精度分别提高3.13%和4.07%，并能有效解决“椒盐”效应及耕地边缘廓线模糊等问题。红边光谱和红边指数的引入分别使随机森林分类器总体精度提高2.39 %和1.63%，并使春、冬小麦的识别能力显著提高，表明红边特征能够帮助分类器更敏感地捕捉不同作物特有的生长特性及物候差异。该研究结果可为机器学习方法及Sentinel 2A卫星在干旱区农业遥感的应用提供参考。

Study of Typical Arid Crops Classification Based on Machine Learning

Accurate and timely crops classification information is of great significance for arid food security monitoring and ecological management. Adding sensitive waveband and improving classification methods are the major development trends of crops classification. In this paper, we carry out crop classification study based on Sentinel 2A time-series remote sensing data, and establish an object-oriented parcel point set in study area, trying to explore the influence of using different classification features on machine learning classification accuracy. Results indicate as follows: (1)Random forest classifier can effectively integrate the benefits of multidimensional vectors such as spectral or vegetation index, all the accuracy of different groups in this study are above 89%, while the supreme overall accuracy up to 94.02%. (2) The classification features extraction method, which was supported by object-oriented parcel point set, can resolve the issue of salt-and-pepper noise and fuzzy parcel boundary well. Meanwhile, it also improves the efficiency and accuracy of machine learning classifier, which can be demonstrated by the result that the classification accuracy of spectral group and index group increased by 3.13% and 4.07% respectively. (3)Red-edge features can help the classifier to capture the phenological differences and unique growth characteristics of different crops. And the introduction of the red-edge spectrum and red-edge index can improve the classification accuracy by 2.39% and 1.63% respectively, while the recognition ability of spring and winter wheat also improved significantly. The result of this study can be referred for the application of the machine learning method and the Sentinel 2A remote sensing data in arid agriculture remote sensing.