基于时序SAR影像的岭南地区耕地内种植草皮识别方法

A Method for Identifying Turf Grass on Cultivated Land in the Lingnan Region Based on SAR Imagery Time-Series

摘要: 占用耕地种植草皮是耕地“非粮化”的表现之一，因此及时、准确掌握耕地内草皮(后简称“草皮”)种植情况对耕地“非粮化”监测具有重要现实意义。针对岭南地区多云多雨复杂观测条件和耕地破碎化严重的问题，本研究以广东省新会区和揭西县为例，提出一种田块尺度的草皮分类方法，首先借助高分辨率光学遥感提取研究区田块数据，然后根据野外调查采集的典型地物样本，构建基于哨兵1号SAR后向散射系数的典型地物时间序列参考曲线，建立田块尺度的包含时序、纹理、统计等信息的特征集，最终构建基于XGBoost机器学习模型的草皮快速识别方法，实现研究区多云多雨条件下草皮的精确识别。结果表明：(1) 草皮后向散射时序特征与水稻等其他地物具有明显区别；(2) 基于时序SAR特征的XGBoost模型验证结果表明，在新会区和揭西县的独立试验查全率和准确率均在80%和93%以上，表明模型具有较好的鲁棒性；(3) 不同特征组合下的控制试验表明，仅使用时序SAR特征可以准确区分光学上易混淆的草皮、其他草地、水稻等地物，使得该模型易向多云多雨的华南地区推广应用。研究结果为耕地“非粮化”精细监测和科学管控、保障粮食安全提供数据支撑和决策支持。

Abstract: The cultivation of turf grass on arable land is one of the manifestations of the “non-grainization” of arable land. Therefore, timely and accurate understanding of the cultivation of turf grass (hereinafter referred to as “turf”) within arable land is of great practical significance for monitoring the “non-grainization” of arable land. In response to the complex observation conditions of the Lingnan region, characterized by frequent cloud cover and rain, as well as the severe fragmentation of arable land, this study takes Xinhui District in Guangdong Province and Jiexi County as examples to propose a parcel-scale turf classification method. First, high-resolution optical remote sensing is used to extract field data in the study area. Then, based on typical ground object samples collected from field surveys, a reference curve for typical ground objects based on the Sentinel-1 SAR backscatter coefficient is constructed. A feature set containing temporal, texture, and statistical information is established at the field scale. Finally, a rapid identification method for turf based on the XGBoost machine learning model is developed, enabling precise identification of turf under the cloudy and rainy conditions of the study area. The results show: (1) the backscatter time series characteristics of turf are significantly different from those of other ground objects such as rice; (2) the validation results of the XGBoost model based on time-series SAR features indicate that the recall and accuracy rates in Xinhui District and Jiexi County are above 80% and 93%, respectively, demonstrating good robustness of the model; (3) control experiments under different feature combinations show that using only time-series SAR features can accurately distinguish between optically confusing turf, other grasslands, rice, and other ground objects, making the model easily applicable to the cloudy and rainy South China region. The research results provide data support and decision-making support for the fine monitoring and scientific management of “non-grainization” of arable land and the safeguarding of food security.