基于时序GF-6与Sentinel-1数据的协同玉米制图指数研究

Synergistic Maize Mapping Index Based on Combination of Time Series GF-6 and Sentinel-1 Data

摘要:

及时准确获取玉米种植分布情况对农业生产管理具有重要意义。针对多云雨区玉米识别中数据缺失与样本依赖难题,本研究提出一种无监督的协同玉米制图指数(Synergistic maize mapping index,SMMI)以实现山西省寿阳县玉米种植区提取。选取2021年5—8月春玉米关键生育期的GF-6 WFV 和Sentinel-1 SAR 影像为数据源,通过耦合GF-6的2个红边波段、SAVI、TVI、RDVI 及Sentinel-1的VH极化特征构建SMMI 进行阈值分类实现春玉米种植区提取。结果表明:利用SMMI 提取春玉米的总体精度达到90. 72% ,玉米用户精度与生产者精度分别达到 88. 07% 与87. 8% ,较未使用雷达特征的指数和基于多特征集的随机森林(Random forest,RF)算法的总体精度分别提升1. 86、0. 95 个百分点,表明SMMI在春玉米识别精度上优于这2种方法,且验证了红边波段与VH极化特征在春玉米识别中的重要性;同时联合光学与雷达特征可实现较单一数据源更优的分类效果。该研究可为多云雨地区的玉米分布制图提供方法参考。

Abstract:

Timely and accurate acquisition of the distribution of maize planting is of great significance to agricultural production management. Aiming at the problems of data loss and sample dependence in maize identification in cloudy and rainy areas, an unsupervised synergistic maize mapping index (SMMI) was proposed to achieve maize extraction in Shouyang County, Shanxi Province. The GF-6 WFV and Sentinel-1 SAR images of the key growth period of spring maize from May to August 2021 were selected as the data sources. By coupling the two red-edge bands of GF-6, SAVI, TVI, RDVI and the VH features of Sentinel-1, SMMI was constructed for threshold classification to extract the distribution of spring maize. The results showed that the overall accuracy of extracting spring maize by using SMMI reached 90. 72% , and the user accuracy and producer accuracy of maize identification reached 88. 07% and 87. 8% , respectively. Compared with the index without using radar features and the multi-feature set-based random forest (RF) algorithm, the overall accuracy was increased by 1. 86, 0. 95 percentage points, respectively. It was indicated that SMMI was significantly superior to the two methods in the recognition accuracy of spring maize. It revealed the importance of the red-edge band and VH polarization characteristics in the identification of spring maize. Combining optical and radar features simultaneously can achieve a better classification effect than a single data source. This research can provide a reference method for mapping the distribution of maize in multi-cloud and rainy areas and had certain potential for agricultural application.