基于无人机和卫星遥感影像的制种玉米田识别纹理特征尺度优选

制种玉米田在高空间分辨率遥感影像上呈现的明显条带状纹理,是有效区分光谱值相近的大田玉米和制种玉米的重要信息.该文在新疆维吾尔自治区奇台县玉米种植区以高空间分辨率的无人机遥感影像为数据源,针对制种玉米识别的纹理特征计算尺度问题,首先采用最近邻内插法对制种玉米和大田玉米样本田块的无人机影像进行重采样,得到不同分辨率的样本;然后用融合Uniform-LBP(local binary pattern)和GLCM(gray level co-occurrence matrix)方法得到提取玉米田块纹理特征合理GLCM参数,其中方向参数为0°、45°、90°和135°这4个方向上的纹理特征值的平均值、距离为5~7像元、灰度级为8;通过多尺度对比分析,得到最适宜区分制种玉米与大田玉米的纹理辨率为0.6~0.9m.最后采用奇台县的0.7m分辨率的Kompsat-3遥感影像进行验证,在多时相EVI(enhanced vegetation index)光谱信息识别玉米的基础上,利用本文确定的纹理分析方法,通过决策树建立规则识别制种玉米,识别精度达90.9%.通过该文的研究,可为高空间分辨率遥感制种玉米田监管提供支撑.

Texture scale analysis and identification of seed maize fields based on UAV and satellite remote sensing images

Abstract: According to the investigation on the spot, it was found that the female parent of seed maize field can be removed tassel and male parent retained the tassel in tasseling stage. Otherwise, the male parent line of seed maize field was cut off and the female parent kept in the mature period. But, grain maize was planted in a uniform pattern. So, seed maize field has the obvious strip texture in high spatial resolution remote sensing images. Which can be used to effectively distinguish the grain maize and seed maize of similar spectral values. In this paper, the high spatial resolution UAV remote sensing image is taken as the data source, and the scaling problem of the texture characteristics in the identification of seed maize is discussed. Firstly, the seed maize and grain maize fields were cut out from the UAV images, and this sample fields were processed by median filtering to remove salt and pepper noise or spots; Next, the seed maize and grain maize fields using nearest neighbor interpolation method to resample and obtain the maize field images with the resolution of 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9 and 1.0 m; Then using the texture extraction and scale analysis method based on Uniform-LBP (Rotation Invariant Uniform Patterns) and GLCM (Gray Level Co-occurrence Matrix), to obtain the rational GLCM values are used as extracting texture features of maize fields and the most appropriate texture resolution scales to distinguish seed maize and grain maize. One, considering the texture feature values of GLCM high redundancy, this paper selected ASM (Angular second moment), Entropy, Contrast and Homogeneity 4 texture feature values which aren''t related to each other in the following research. Two, because the same texture feature values of the same field affected by the texture analysis direction of GLCM, the paper use Uniform-LBP on maize sample images to obtain rotation invariant LBP image. Experiments showed that four texture feature values of maize fields will be a little fluctuation with the change of direction after Uniform-LBP treatment, but the overall amplitude is smaller, so in order to eliminate the influence of parameters of the direction, in this paper, the direction parameters is the average value of the texture features of the four directions for 0°, 45°, 90° and 135°. Three, it is found that distance is from 4 to 10 pixels, GLCM texture feature values tends to be stable, particularly, when the distance parameter is from 5 pixels to 7 pixels, which satisfies the distribution of seed maize stripe texture. According to seed maize in the study area was planted by the ratio of male to female from 1:6 to 1:8, line spacing is from 0.6m to 0.8m, so the strip texture spacing is from 3.6m to 4.8m under 0.7m resolution. Four, results showed the texture characteristic values of maize are not affected by gray level compression, so the gray level parameter choose 8 to reduce the amount of computation. Five, it is found when resolution from 0.6 to 0.9 m, texture feature values differ greatly and it is easy to distinguish seed maize and grain maize, so the most appropriate texture resolution scales is from 0.6m to 0.9m. Finally, maize planting area in Qitai Country, Xinjiang Uygur Autonomous Region was take as the study area to verify, using KOMPSAT-3 image of 0.7 m resolution, based on maize recognition results with multi-temporal EVI spectral information, using the texture analysis method in this paper, combined with the rules established by decision tree, to recognize the seed maize. The results show that seed maize identification precision reached 90.9% at 0.7m resolution, basically meeting the needs of seed maize identification requirements, which can provide support for the high spatial resolution remote sensing seed maize field fine supervision