应用Sentinel-2A NDVI时间序列和面向对象决策树方法的农作物分类

农作物种植结构是农业生产活动对土地利用的表现形式。及时精确地获取农作物的空间分布信息对指导农业生产、合理分配资源以及解决粮食安全问题等具有重要意义。目前农作物信息提取研究大多局限于中低分辨率遥感影像的NDVI时间序列,影响了作物空间分布信息提取的准确性。随着Sentinel-2A卫星成功发射,为高分辨率NDVI时间序列的构建提供了可能。本文以黑龙江省北安市为研究区,基于覆盖完整生育期的Sentinel-2A多光谱数据,构建10 m分辨率的NDVI时间序列数据集,利用 Savitzky Golay (S-G) 滤波器对 Sentinel-2A NDVI时间序列数据进行平滑。基于典型时相的多光谱数据和NDVI时间序列构建面向对象决策树分类模型进行作物类型遥感识别。通过对样本的NDVI时间序列曲线分析,可以得出NDVI时间序列能够清晰地区分作物物候差异。此外,本文还利用面向对象分类和支持向量机（Support Vector Machine, SVM）分类两种方法,对典型时相的多光谱数据进行了作物分类对比实验,并对结果进行了对比分析。研究结果表明：① 典型时相多光谱数据引入平滑重构后的NDVI时间序列能够更好地描述作物的物候特性,能够准确刻画研究区作物发育情况,有效区分各类作物;② 通过对比分类实验发现,典型时相多光谱数据引入NDVI时间序列特征,增强了不同作物之间的光谱差异,提高了作物分类精度,总体精度和kappa系数较典型时相多光谱数据进行分类的结果分别提高了7.7% 和0.055;③ 基于面向对象的决策树分类模型在作物分类的结果中精度最高,总体精度为96.2%,kappa系数为0.892。本研究的方法为其他大区域农作物的分类提供了重要参考和借鉴价值。

Crop Mapping based on Sentinel-2A NDVI Time Series Using Object-Oriented Classification and Decision Tree Model

Cropping pattern is a fundamental aspect of land use. Obtaining accurate and timely crop spatial distribution information is very important to guide agricultural production, rational allocation of resources, and help solve the problem of food security. NDVI (Normalized Difference Vegetation Index) time series have been widely used in collecting vegetation information. Identification and information extraction of different crops can be effectively achieved by analyzing the growth period of crops and the time-series spectrum characteristics of NDVI. Most existing studies on NDVI time series are limited to moderate or low resolution remote sensing imagery, which affects the accuracy of extracting crop spatial distribution information. With the successful launch of the satellite Sentinel-2A, more opportunities have emerged for the construction of NDVI time series with high temporal and spatial resolution. In this paper, by use of typical phase Sentinel-2A imagery for Beian city, a NDVI dataset with a spatial resolution of 10 m covering the whole growth period of April-October was generated based on the Savitaky-Golay filtering smoothing method, and crop identification was implemented based on decision tree model and object-oriented classification. By analyzing NDVI time series curves of crop samples, we concluded that NDVI time series was able to clearly distinguish crop phenological differences and capture crop specific features in the study area. Furthermore, we also discussed the classification accuracy based on the typical phase data by the methods of object-oriented classification and support vector machine. Taking the field sample survey datum as true value, we analyzed the results of the two classification methods. The results show that the processed NDVI time series with high resolution over the entire crop growth cycle represent different crop phenological characteristics appropriately. It is able to reflect the crops growth condition accurately and distinguish different crops effectively. The decision tree model integrated with the object-oriented classification method had the highest accuracy in crop classification as compared to other classification methods, with its overall accuracy and kappa coefficient being 96.2% and 0.892, respectively. This research show that Sentinel-2A NDVI with high resolution can be used for crop mapping, and can be applied to crop classification over large areas, thanks to Sentinel-2A imagery's wide coverage. Furthermore, the Savitzky Golay (S-G) method was used for NDVI time series smoothing, and the results indicate that the processed NDVI time series can better represent crop phenological characteristics. Then the decision tree model integrated with the object-oriented classification method was used for crop classification based on typical phase multi-spectral imagery and the smoothed NDVI time series, which improved the overall accuracy by 7.7% and the kappa coefficient by 0.055. The approach proposed in this paper provides important reference for crop mapping in other agricultural regions.