地块尺度下基于多源卫星遥感数据的粮食作物识别

开展粮食作物监测对于国家粮食安全具有重要意义。在传统像元尺度下，利用单一遥感数据进行粮食作物监测，识别精度往往较低，提取的作物地块破碎，难以满足应用需求。为此，该研究以山东省青岛市黄岛区为研究区，提出了一套地块尺度下综合多源卫星遥感数据（包括高分辨率数据、多时相数据、高光谱数据）与土地利用调查矢量数据的粮食作物信息识别方法。首先，对高分辨率数据进行分割获取耕地地块矢量数据；其次，基于多源卫星遥感数据提取地块级时空谱特征；再次，利用样本数据计算特征类间可分性，并进行特征优选；最后，构建基于二次多项式支持向量机的主要粮食作物（春玉米）识别方法。结果表明：1）该研究所提的方法可以有效进行粮食作物信息识别，基于地块数统计的识别精度为89.7%；2）利用光谱特征、植被指数、纹理特征组合得到的识别结果精度最优，基于像元数统计的精度为97.1%，与传统方法相比提高了24.2个百分点，且提取的地块信息更完整。该研究成果可支持粮食作物种植用地的调查与监测，也可为耕地非粮化时空演变与分析提供新的思路。

Identification of grain crop using multi-source satellite remote sensing data at field parcel scale

Abstract: Grain crops are of great significance to the national food security in the world in recent years. Among them, the single remote sensing data source can usually be used to monitor the distribution of grain crops at the pixel scale. However, the monitoring data of grain crop can be constantly contaminated with the serious pepper noise. The current accuracy cannot fully meet the harsh requirements of cropland management in smart agriculture. In this study, a modified strategy was proposed to identify the grain crop information at the field parcel scale using the multi-source satellite data (including high-resolution data (HR), multi-temporal data (MT), and hyperspectral data (HS)), and vector data of land-use type. A study area was selected as the 13 km×12 km complex planting region in the Huangdao District of Qingdao City, Shandong Province, China. The data sources were collected from the ZY1-02D Visible Near-Infrared Camera (VNIC), Advanced Hyperspectral Imager (AHSI), and Landsat-8 Operational Land Imager (OLI). The experiment was performed on the following steps. Firstly, the vector data was used to obtain the cultivated and non-cultivated land boundaries, where the non-cultivated land was masked in the study area. Secondly, the watershed algorithm was utilized to segment the HR for the boundaries of homogeneous crop field parcels. Thirdly, the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) were calculated to determine the thresholds of vegetated field parcels. After that, the spatiotemporal spectrum feature datasets were constructed at the field parcel scale using multi-source satellite remote sensing data, including the spectral features (SFs), vegetation indices (VIs), and texture features (TFs). Then, an independent sample t-test method was adopted to calculate the separability between the grain crops and non-grain crops within different feature types. The optimal subset of features within each feature type was determined to train the classifier with the diagonal covariance matrix as the discriminant analysis. Seven sets of features were constructed to evaluate the different combinations of the optimal subset of each feature type. A quadratic polynomial support vector machine (QSVM) identification was then carried out to evaluate the accuracy of the system. Subsequently, the contribution of each feature type to the accuracy was analyzed under the different field parcel sizes. Moreover, the optimal feature set was achieved to compare the identification accuracy. Finally, the better accuracy was determined at the field parcel scale and the traditional pixel scale under the optimal feature set. The results showed that: 1) The proposed strategy performed better to acquire the grain crop distribution of cultivated land, with an identification accuracy of 89.7% using the number of field parcels. 2) The maximum identification accuracy of 97.1% was achieved at the field parcel scale using the number of pixels, in terms of the optimal feature set with the SFs, VIs, and TFs. The accuracy was improved by about 24.2%, compared with the traditional pixel scale. There was also much more complete crop field parcels to be identified. 3) The HR_VI and HS_SF can be expected to significantly improve the identification accuracy of small and large field parcels, respectively. In medium-sized field parcels, both HR_VI and HR_TF were contributed to the high identification accuracy of cultivated land change. The finding can also provide a strong reference to efficiently utilize the cultivated land.