基于时空滤波Sentinel-1时序数据的田块尺度岭南作物分布提取

为解决岭南地区作物制图中光学影像时空覆盖缺失以及作物种植结构复杂导致不确定性高等问题，以满足对高精度作物制图日益迫切的应用需求。该研究提出了一种基于Sentinel-1的双极化时间序列和作物物候信息建立时空维度典型物候特征的作物分类方法，选择广州市南沙区为试验区，通过XGBoost机器学习算法，实现了田块尺度下的不同作物类型的准确识别及种植面积的精细提取。结果表明：1）相比像素时序特征分类结果，经过雷达时空滤波后的田块时序特征分类方法不仅能有效抑制合成孔径雷达（Synthetic Aperture Radar，SAR）影像斑点噪声产生，而且总体分类精度和Kappa系数分别提高了12.5个百分点、0.19；2）与仅基于Sentinel-1(VV+VH)时空滤波后的全年时序特征分类方法相比，在分类过程中添加物候特征变量的方法表现出更高的精度，Kappa系数可达0.91，甘蔗和香蕉播种面积提取精度分别达到82.04%和71.01%。研究结果表明中高分辨率SAR影像（10 m×10 m）时间序列结合XGBoost算法和雷达数据时空滤波策略可实现区域作物准确识别及种植面积提取，同时，可从数据源与分类方法上为岭南地区农业遥感应用提供思路，对该地区农业灾害管理和灾后救助具有重要参考价值。

Extracting field-scale crop distribution in Lingnan using spatiotemporal filtering of Sentinel-1 time-series data

Timely and effective estimation has been critical to capturing the distribution information of field-level crops in precision agriculture. However, the spatiotemporal coverage of optical images can be confined to the fragmentation and heterogeneity of crops in cloudy and rainy south of China. In this study, a rapid and accurate extraction was implemented for the field-scale crop distribution in the Lingnan region of China using spatiotemporal filtering of Sentinel-1 time-series data and the crop phenological information with the field data. The European Space Agency Sentinel-1A (S1A) satellite data was selected to effectively alleviate the insufficient optical image available in the study area, particularly with the spatial resolution of 10 m and the 12-day revisit period. A field-scale classification was also proposed using the Sentinel-1 dual-polarization time-series and crop phenological information. Specifically, the typical phenological characteristics were established in the spatiotemporal dimensions. A field experiment was performed on the XGBoost machine learning for the diverse plant types in Nansha District, Guangzhou City in the study area. As such, the high-precision mapping of crop type was achieved using the time-series of the Sentinel-1 data. A filtering approach was proposed to suppress the spot noise from two levels in the spatiotemporal dimensions. The mixed cell noise at the edge was effectively suppressed to smooth the abnormal fluctuations in the time-series, compared with the traditional filtering in the classification of synthetic aperture radar data. Firstly, the cropland area in each field was vectorized using GF-2 images. The field sizes were then replaced with the fixed filter window for the spatiotemporal denoising of SAR data before classification, which was used to obtain higher classification accuracy. Secondly, the Sentinel-1 dual-polarization (VV + VH) images along with the ground sample data were utilized to train and evaluate the performance of the field- and pixel-scale extraction in the crop type mapping. Lastly, the phenological characteristic variables were constructed by the time-series features, where the field-scale extraction was combined to improve the accuracy of field-scale mapping. The results showed as follows. 1) The classification using the time-series features of the field effectively suppressed the speckle noises in the SAR images, where the overall accuracy and the Kappa coefficient were improved by 12.5 percentage points and 0.19, respectively, compared with the time-series features of the pixel. 2) Compared with the classification using the only time-series features of Sentinel-1 (VV+VH) after spatiotemporal filtering, the phenological feature variables in the classification presented the better accuracy, where the Kappa coefficient was 0.91, while the sown area accuracy of sugarcane and banana reached 82.04% and 71.01%, respectively. Consequently, the time-series of Sentinel-1 image combined with the XGBoost and the radar data spatiotemporal filtering can be widely expected to achieve highly accurate crop identification and planting area extraction. At the same time, the finding can provide a strong reference for agricultural remote sensing from the data source and classification, disaster management, and post-disaster relief in the Lingnan region.