基于主被动遥感数据和面向对象的大蒜识别

针对开封市大蒜种植破碎化程度高，光学数据难以高精度、快速提取问题。该研究基于谷歌地球引擎（Google Earth Engine，GEE）云平台、随机森林算法（Random Forest，RF）和面向对象方法，选择融合Sentinel-1卫星的后向散射系数与Sentinel-2卫星的光谱、光谱指数及纹理特征，分别应用10 m与加入植被红边波段的20 m空间分辨率遥感数据，探究不同特征组合对改善大蒜识别精度的性能。试验结果表明：应用10 m空间分辨率的Sentinel主被动遥感数据，在简单非迭代聚类（Simple Non-iterative Clustering，SNIC）分割尺度为5，灰度共生矩阵（Gray-level Co-occurrence Matrix，GLCM）邻域值为4，7个纹理特征选择第一、二主成分时，分类总体精度和Kappa系数最高，为94.54%、0.93，大蒜的制图精度和用户精度为97.83%、96.38%。应用加入植被红边波段的20m空间分辨率Sentinel主被动遥感数据，在SNIC分割尺度为3，GLCM邻域值为4，7个纹理特征选择第一、二主成分时，分类总体精度和Kappa系数最高，为94.14%、0.92，大蒜的制图精度和用户精度为95.72%、98.81%。单独使用Sentinel-2光学数据，加入植被红边波段的20m分辨率数据相对10 m分辨率数据，大蒜制图精度和用户精度分别提高0.49%和4.38%。单独使用时序Sentinel-1 SAR数据，10 m空间分辨率数据的大蒜制图精度和用户精度优于20 m分辨率数据0.66%和2.07%。本研究为遥感数据识别生长周期相同或重叠的大宗、小宗经济作物提供技术参考。

Identification of garlic based on active and passive remote sensing data and object-oriented technology

Abstract: Garlic as an important cash crop was widely planted in Henan, Shandong province and other areas. In the past few years, the price of garlic had fluctuated frequently and seriously affected farmers'' decision and enthusiasm of garlic plantation, which was not conducive to the healthy and sustainable development of the garlic market. Garlic crop planting was often mixed with other ground crops, making it difficult to quickly extract garlic with high precision based on optical remote sensing data. In this research, Kaifeng of Henan province with complex planting structure was selected as a study area. With the aim to further improve the identification accuracy and efficiency of fragmented garlic by using remote sensing data, an object-oriented method based on Google Earth Engine (GEE) platform and random forest (RF) algorithm was used by integrating the active and passive remote sensing data of Sentinel satellites. The normalized difference vegetation index (NDVI) time series data can reflect the phenological changes of the garlic and other crops, which can be used to improve the separability of garlic and other land cover types. Based on the maximum difference of NDVI and the features of RF algorithm, this paper selected the monthly mean Sentinel-2 optical data in March 2021. And the time series Sentinel-1 synthetic aperture radar (SAR) backscattering coefficients of monthly mean data was selected from November 2020 to May 2021. It included two steps before classification. The first one was to use the simple non-iterative clustering (SNIC) for integrating data segmentation and selected the best segmentation scale with the highest classification overall accuracy and kappa coefficient. The second one was to use the gray-level co-occurrence matrix (GLCM) algorithm selecting the best one within three neighborhood values (4,8,16) to calculate the seven texture features of synthetic optical data and use the principal components analysis (PCA) algorithm to reduce those data dimensions. Through integrating Sentinel passive and active remote sensing data of 10 m spatial resolution or 20 m spatial resolution with incorporating the vegetation red edge bands to explore the identify accuracy improvement of garlic on combining various features groups of spectral features, backscattering coefficients, vegetation index features and different principal component groups of texture features, the result showed that using the 10 m spatial resolution active and passive Sentinel remote sensing data, the highest classification overall accuracy and Kappa coefficient reached 94.54%, 0.93, and the producer''s and user''s accuracy of garlic were 97.83% and 96.38% at the SNIC segmentation scale of 5, GLCM neighborhood value of 4 and the first and second principal components of the 7 texture features. Using the 20 m spatial resolution active and passive Sentinel remote sensing data with three vegetation red edge bands, the highest classification overall accuracy and Kappa coefficient reached 94.14%, 0.92, and the producer''s and user''s accuracy of garlic were 95.72% and 98.81% at the SNIC segmentation scale of 3, GLCM neighborhood value of 4 and the first and second principal components of the 7 texture features. The producer''s and user''s accuracy of 20m spatial resolution Sentinel-2 data with three vegetation red edge bands were improved 0.49% and 4.38% compared with 10m spatial resolution data. The producer''s and user''s accuracy of 10 m resolution time series Sentinel-1 SAR data were improved 0.66% and 2.07% compared with 20m spatial resolution time series SAR data. Integrating Sentinel active and passive remote sensing data, which fully used the spectral and structural information, the overall accuracy and Kappa coefficient were higher than alone using the optical or time series SAR data. Moreover, the classification overall accuracy and Kappa coefficient was highest with integrating Sentinel passive and active remote sensing data at 10 m high spatial resolution. Therefore, the SNIC, GLCM, PCA, RF algorithm and GEE platform used in this paper had a good promotion value for accurate and efficient obtaining garlic planting area, especially for using GF satellites data to obtain the Four-Huaiqing Chinese medicine area.