基于Radarsat-2影像的复杂种植结构下旱地作物识别

为提高基于Radarsat-2旱地作物识别的精度，该文研究了一种复杂种植结构背景下具有共同生长期作物的识别方法。研究区为一个12 km×12 km的样方，位于内蒙古上库力农场额很队，以春小麦、油菜2种共同生长期作物为识别对象，利用Spot-6影像和Radarsat-2影像，在数据预处理的基础上分析研究区内典型地物样本的后向散射系数在不同极化波段上的变化特征，根据该变化特征设计图像增强算法，然后基于图像增强后的影像设定合理的阈值实现作物识别提取。结果表明：该方法准确识别并有效提取了共同生长期作物春小麦和油菜的种植面积，总体精度达到97%，Kappa系数为0.96。该方法简便、快捷、可靠，为春小麦、油菜等旱地共同生长期作物种植面积提取提供重要的科学技术支撑。

Dryland crops recognition under complex planting structure based on Radarsat-2 images

Abstract: Acreage information of crop planting is one of important scientific bases for making national food policies and economic plans. Therefore, it is very important to carry out the study about crop recognition and monitoring. Optical images are not always available in the key growth period of crops, owing to the cloudy and rainy weather. Synthetic aperture radar (SAR) remote sensing has the advantages of all-weather, all-time, high resolution and wide coverage. However, there are few researches about extraction of dryland crops that have the same growing season using SAR data, and the extraction accuracy of dryland crops is not high (usually under 90%). In order to improve the accuracy of dryland crops recognition based on Radarsat-2, a method of dryland crops extraction was proposed in this paper. The study area was a sample that was part of Shangkuli Farm in Inner Mongolia Autonomous Region and the area of the sample was 12 km×12 km. The objects included spring wheat, oilseed rape, forest, grassland and water in the study area. Research data included one scene of SPOT-6 image and 2 scenes of Radarsat-2 images. These data were disposed and analyzed using NEST software and ENVI software. Data pretreatment process included radiation correction, geometric correction, producing backscatter coefficient, range-doppler terrain correction, projection transformation, and subset images. The backscatter coefficient variation characteristics of the typical surface features in different polarization band were analyzed based on samples, and images enhancement algorithm was designed according to the variation characteristics. Radarsat-2 images on August 3 were analyzed and result showed that those images included 5 kinds of typical surface features: water bodies, forests, grasslands, oilseed rape and spring wheat. Backscatter coefficient of water was minimum among the 3 kinds of polarization (they were HH, VV and HV), and that of oilseed rape was maximum among the 3 kinds of polarization i.e. HH, VV and HV (the first and second letter represent the polarization way of launching and receiving electromagnetic waves by radar respectively; H and V mean horizontal and vertical polarization respectively), so the function of the image enhancement algorithm of HH+HV+VV would enhance the information of oilseed rape and water on images. Through experimental analysis based on samples, it was found that backscatter coefficient was -84.01--60.52 dB for water, more than -36 dB for oilseed rape, and more than -49.68 dB for other 2 classifications, so a pix was defined as oilseed rape if the value of the pix was more than -36. The maximum value of the backscatter coefficient for water was -60.52 dB and the minimum value of the backscatter coefficient for other 3 kinds of surface features was -49.68 dB, so a median value -55 was set as the threshold that was used to extract water information. And a pix was defined as water if the value of this pix was more than -55. Thus water and oilseed rape classification layer were gained. Information of spring wheat was prominent on the Radarsat-2 images that were gained on Jul 10th, and the backscatter coefficient of spring wheat on HH polarization was more than that on HV and VV polarization, and other surface features didn't have this characteristic. So the function of the image enhancement algorithm of HH-(HV+VV) would enhance the information of spring wheat on images, and sample analysis results showed that the backscatter coefficient of spring wheat was more than 22.6 dB on the HH-(HV+VV) band. A pix was defined as spring wheat if the value of this pix was more than 22.6 and spring wheat classification layer was gained. For water, oilseed rape and spring wheat classification layers, clustering processing was conducted with 9×9 window size and median filtering method was used to suppress noise with 9×9 kernel size. Crops were harvested when gaining the SPOT-6 image, but the forest and grass were growing, so the spectral characteristics were different between arable land and forest and grass land. Arable land information was gained based on SPOT-6 image, and spring wheat and oilseed rape were divided based on the extracted crop results. The spring wheat, oilseed rape and arable land classification layer were combined, and integrated figure was made based on the decision tree classification method. Results showed that the planting acreage of spring wheat and oilseed rape were effectively extracted using this method, the overall accuracy was 97%, and the Kappa coefficient was 0.96. This method was simple, fast and reliable, and provided an important scientific and technical support for the extraction of planting acreage of spring wheat and oilseed rape.