基于多时相双极化SAR数据的作物种植面积提取

及时准确地获取农作物的空间分布信息和种植面积,在农业生产管理与农业政策的制定等斱面具有非常重要的作用。本文以多时相Sentinel-1A影像(4月17日、5月5日、6月16日、7月22日、8月27日、9月2日)为主要数据源,根据研究区作物的物候特征,提取棉花、玉米和果树在不同生长期的后向散射系数(Sigma)和归一化后向散射系数(Gamma)。通过对作物不同极化、不同时相后向散射系数的统计,建立散射特征时序变化曲线,幵分析其特征。利用人工神经网络(Artificial neural network)、支持向量机(Support vector machine)和随机森林(Random forest) 3种分类斱法对研究区的主要农作物迚行分类识别以及种植面积提取,幵对分类结果对比分析和验证。结果表明, 1)棉花的后向散射系数在6月现蕾期和7月开花期明显上升,8月仹达最高值,变化特征最明显,易与其他作物区分;玉米和果树的后向散射系数在9月仹与其他地物之间表现出显著差异。2)相较于神经网络和支持向量机,随机森林的分类效果最好,总体精度达88.97%。其中,对棉花和果园的分类精度为90.88%和93.17%,对玉米的分类效果最差,仅有71.6%。综上所述,多时相双极化SAR数据在不同类型作物的识别及面积提取斱面具有一定的应用潜力。

Extraction of crop acreage based on multi-temporal and dual-polarization SAR data

It plays a vital role in agricultural production management and agricultural policy formulation to acquire spatial distribution information and planting area of crops timely and accurately. In this paper, according to the phenological characteristics of crops, the back scattering coefficients (Sigma) and normalized back scattering coefficients (Gamma) of cotton, corn and orchard in different growth periods were extracted by using multi-temporal Sentinel-1A images (April 17, May 5, June 16, July 22, August 27, and September 2). The time-series change curves of the scattering characteristics were established and the characteristics were analyzed by using backscattering coefficients of crops with different polarizations and different time phases. Artificial neural network (ANN), support vector machine (SVM), and random forest (RF) were used to classify and identify cotton, corn and orchard. On this basis, the classification results were compared and analyzed, and the accuracy was verified. The backscattering coefficient of cotton increased significantly in June and July with the highest value in August, of which the changes were most obvious and easily distinguished from other crops. The backscattering coefficients of corn and fruit trees were significantly different from those of other land objects in September. The optimal classification was obtained by using random forest (the overall accuracy was up to 88.97%) than by using neural network and support vector machine. The classification accuracy for cotton and orchard was 90.88% and 93.17%, and the classification effect of corn is the worst, only 71.6%. In conclusion, multi-temporal and dual-polarization SAR data has certain application potential in the identification and area extraction of different crops.